I spent most of last week in San Francisco, but mostly sat in my hotel obsessing over my Stanford talk while trying to work out where I’d find the specific colour presents my kids had requested. Arabella wanted purple. Max wanted green.

Jenni is the main reason I went to SF. She invited me to Stanford and it was such a privilege. Plus, she’s brilliant. Sharp, high-agency, but warm and very personable. She also runs an AI consultancy for CEOs and gets my highest recommendation. If you get the chance to work with her, take it.

SF has this energy - how do I go faster, grow bigger, become more successful?

I get it. I like that energy. But I also came home thinking that my own version of ambition looks a bit different.

Yesterday I finally read the profile of Lenny from Lenny’s Newsletter and a lot of it stuck with me. Because it was about building a life where you can care deeply about the work without letting the work eat everything else.

That felt so ‘me’ I had to text him and let him know!

I spent an unreasonable amount of time on my Stanford talk. Not because I didn’t know what I wanted to say, but because I cared about how it landed and was presented. And also because I was trying to make the course and the talk at the same time.

Working on it also made it clearer to me that I sit in this weird middle ground where non-technical people think I’m technical, and developers don’t. But I think a lot of you are somewhere in that messy middle too. Truth is, it’s all a bit messy - everyone is still figuring out agents and AI.

Working with agents isn’t about becoming a developer, but understanding the shape of things; files, tools, systems. With a thick dollop of taste on top.

If you can steer agents, you can get technical pretty quickly. I did. You just have to be realistic that you’ll hit tons of bumps in the road. Your job is to use your agent to figure it out - and you’ll probably learn something that’ll come up again and again.

That’s what I want Ben’s Bites to be more about. Taking you along my exploration: what I’m seeing, what I’m trying, how I’m thinking about it, what bumps I ran into, come along and try whatever sounds useful. Too much ‘education’ out there is just thirsty growth-hacks to sell you something.

I’ve sold a company. I’ve got three tiny kids at home. I want to do excellent work, make good money, back great companies, and build something useful without accidentally creating a job I don’t want.

There are big, valuable companies that want me to work with them. They’d give me clout, access, money. But I hesitate because I don’t know if I can give that kind of thing my all.

Even with my fund. I’m invested in funds who have more money, better process, bigger pipelines. And yet my funds are outperforming.

But I struggle to fundraise because my story isn’t presented cleanly and my process doesn’t look like a workflow. I tinker. I talk to developers. I try tools. I see what people click. I back founders building things I think will matter.

It’s all connected. The newsletter gives me a read on what builders care about. The fund lets me back the tools that might become important. The course/workshop stuff is me trying to teach the shift I’m living through myself.

Devtools built for developers today become the tools agents use. Humans steering, agents operating. They’ll pick up the tools, compose them, run the commands, change the files, connect the systems.

We just gotta understand enough of the shape of the work to direct it well.

A lot of what I do comes down to feel. Caring about the thing enough to make it good, but not needing to turn every good thing into a machine. And squeezing every last drop for growth’s sake.

So I came back from SF without a plan to scale. Mostly I came back thinking, I want to keep building for this new class of builders. People who are curious, increasingly technical, and trying to use AI to become more capable.

That feels like a good place to spend my time. Raising the floor, not the ceiling (h/t Jenni/Jen’s Bites).

Exploring, tinkering and teaching.

p.s. the kids got their toys, and I’ll continue working on the ‘course’ with care 😊

p.p.s. speaking of care, my brother Adam just launched Hono UI - a UI kit (like Shadcn) but for projects that use Hono. Proud of him! Do me a favour and blow up his launch post 🙏

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there's more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• OpenAI released GPT-5.5 - On paper, it’s a much better model than Opus 4.7 (minus frontend design - but using their new image gen model and asking 5.5 to turn that into code is a great workaround). This comes with a price increase making it 2x more expensive than GPT-5.4 and even slightly more than Opus 4.7 on a per-token basis. But claims it’s 40% more token efficient, so per-task cost doesn't change much (Ramp reported similar results).

  • It’s become my default model since Anthropic cracked down on usage outside of their own harness. I’m in ‘thinking:low’ camp, super smart and fast.

  • I’m also trying to use the codex desktop app over a terminal and there’s lots to like - but I’m just desperate for cloud/mobile use too.

  • One more thing… it’s clear OpenAI is focusing on builders, they’re shipping like crazy, are very vocal on X, responding to users and resetting rate limits. We love to see it.

• Claude got worse. It’s official. Anthropic posted a breakdown of what and why it happened. Basically, changing the default thinking mode and system prompt changes led to a combined fall in the quality of Claude's responses, especially in Claude Code. They still clarify that they don’t switch to a quantised or “worse” model.

  • Memory on Claude Managed Agents is now in public beta. Context: Managed Agents is Claude’s API product for companies to easily plug in claude code-like agents into their products without managing the hassle of infrastructure.

• Cursor needs compute and a model. XAI needs users. And competition is higher than ever for coding agents. Cursor was raising $2Bn at $50Bn but has been paused because they struck a deal with SpaceX/XAI that gives SpaceX the option to buy Cursor for $60B later in 2026, or instead pay $10B for the collaboration. It’s a weird structure I’ve not seen before - so we will see what happens. Never a dull day.

• Your CRM knows nothing because you never update it. Lightfield captures every email, call, and meeting automatically — then lets you ask anything. “Why do we keep losing to [competitor]?” Real answers from real conversations. Get 3 months off with code BENSBITESS13.*

Subscribe now

My feed

• QuickCompare helps you choose the right LLM for your use case. Compare 50+ models on your own data in minutes with Trismik. Try free now!*

• Flipbook - every pixel on your screen, streamed live directly from a model. Another similar demo.

• Clicky lets you talk to AI and spawn an agent on your Mac. Farza built and open-sourced a version of this a few days ago, but the new clicky is closed source. OpenAI already had a version of this which you can clone.

• Things I learned working at OpenAI.

• AI isn’t replacing anyone.

• Exa for Claude - plugin to give your Claude access to websites, people, companies, and more.

• Ora.run - scan and rank how well agents can find and use your business.

• Tolaria is another app to read/write markdown files on your desktop with second-brain-esque features.

• Reelful - from a camera roll to a finished reel in just 10 minutes.

• trunks by layerbrain - turn your own storage into git remote with a minimal CLI.

• webpull - instantly pull any website into a directory of clean, searchable markdown files.

• slacrawl - cli terminal app for slack with sqlite backend.

• When and how to use MCP for apps built for production.

• create-agent-tui - skill for building your own agent harness + terminal UI.

• Aman (ex-Harvey) built an offline AI app to track his skin’s response to different diets and products with local models. (testflight) (how he built it)

• OpenAI is updating its partnership with Microsoft. Separately, Sam and Greg did an interview recently (their first one together?)

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

For a few months, it felt like Google had won the image generation space. But OpenAI is back in the game. ChatGPT Images 2.0 is miles ahead of anything. It’s beyond impressive at text, I haven’t seen any generation with typos, even with hundreds of words per image. See this example I created:

It’s also really good at creating realistic pictures, like this one of Professor Ben.

Oh, sorry, that one’s real. Ben was at Stanford this Wednesday, teaching how to build with AI agents.

Image generation is also available in the Codex app as a skill. Use it with thinking models to get the best results—that lets it think and use code/tool calls (like creating a QR from a link, searching logos from the web) and then use them as reference images. It can also create images, reflect on them and improve the generation.

People are creating realistic UI screenshots, multi-page illustrated magazines, personal style recommendations and creative QR codes using the new model.

The “generate UI as image” bit is interesting. Maybe there’s finally a solution to GPT-5.4’s lack of design taste. The latest coding models are fairly good at turning screenshots into code, but there are still gaps.

Last weekend, I tested a bunch of tools/models on implementing a design (for an ads storefront for Ben’s Bites) from a screenshot. I found:

• Claude Design > Magicpath AI > Raw models (like Gemini 3.1 Pro/Opus 4.6 in their web apps), when it comes to understanding the concept and making something usable, not just copying the pixel-by-pixel look (ironically, Gemini won that).

• When asked to turn designs into a real working app, there was a major drift in how the apps looked. Opus 4.7 did better than GPT-5.4 at visually matching the reference screenshot. Though GPT-5.4’s code was more functional, and the unseen pages (like the admin panel) had a consistent design with the rest of the app.

Also, in many cases, the assets (hero image, icons, background textures) make the UI in a “generated image” stand out. When replicating that UI from a screenshot, you get the barebones UI with the correct buttons and the layout, but without those assets, and the output falls short of expectations.

Ben’s Bites is brought to you by TinyFish

Funny how AI agents can write entire apps but can’t work on the live web. Playwright scripts break, raw fetches eat your context, bot detection blocks you, nothing’s scalable. TinyFish gives search, fetch, stealth browser, web agent, all managed in one API. Try it free. Comes with a CLI + Skill.

Headlines

• OpenAI has a new product for Business, Enterprise and Edu users - Workspace Agents. Codex-powered agents inside ChatGPT with a persona, task and access to external tools (like Linear) and accessible for Slack as well. These agents will also replace custom GPTs down the line (finally). Read more.

• Gemini Deep Research API now offers two configurations based on 3.1 Pro. It claims the best performance in web research and finding hard facts. Plus, it gets MCP support and can create charts using Nano Banana or HTML.

• Cursor and SpaceX are working together - Cursor will train coding models on SpaceX’s GPUs and likely share them with xAI. SpaceX can, in turn, acquire Cursor later this year for $60B, or pay $10B for the partnership if it doesn’t. On a similar note, Thinking Machines also just signed a multi-billion-dollar Google Cloud deal.

• Give your Droid a computer - You can now give your Droid an always-on machine with its own filesystem, credentials, and config for it to keep working on your tasks. This can be in the cloud (managed by Factory), or you can bring your own device.

Subscribe now

My feed

• Chronicle - Cursor for slides. Never build a deck from scratch again. Turn ideas into stunning presentations in minutes.*

• ChatGPT for Excel and Google Sheets are now in beta - build new sheets, fix formulas, explain models, and update workbooks in place. (read more)

• /ultrareview in Claude Code (research preview) lets you run bug-hunting agents in the cloud before merging riskier changes like auth, data migrations, or other critical code paths.

• OpenAI built an open-source viewer for chat data and Codex session logs - Euphony.

• Sierra is piloting an AI-native interview - debugging/review focused interviews where candidates improve a medium-sized codebase with coding agents.

• ml-intern from Hugging Face - open-source research agent to come up with experiments, and run them.

• Clawputer - Managed OpenClaw agent inside an always-on sandbox.

• Kami - design skill for AI-native docs, resumes, portfolios, long docs, and slides.

• noscroll - an AI that doomscrolls X for you and texts you just the signal. In my experience, this is easy to claim and hard to get right.

• Monologue has a new Notes feature for thinking out loud when you don’t know the exact words you want to dictate.

• Fin is moving beyond customer support into sales - using the same business context and integrations to qualify leads and book meetings.

• Perplexity post-trained a Qwen-based model to handle search and tool calls for cheaper, and it’s already serving a meaningful chunk of traffic.

• The next Slack won’t look like Slack, and Ando looks like one early attempt at that.

• Frontend in 2026 - for and against the frameworks and abstractions dominant today.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ve been playing a lot with Claude Cowork for my talk at Stanford later today. It’s comically bad for the average jane.

A lot of capabilities are enabled using connectors and plugins, but if you don’t know that, good luck getting it to do anything. Can’t seem to send emails, install a skill or tell me about what potential connectors are there.

Scheduled tasks in Cowork stop when you shut the lid, but Routines (similar thing) in Claude Code do not. Cowork is just now getting Artifacts - the mini apps on the Claude chat app that started the vibe-coding wave.

I can search for all of this, yes. but an average user will not. and they’ll walk away thinking AI is hype for the next 6 months/a year.

Anyway, the Cerebral Valley AI Summit is coming back to London on June 24th. I’m planning to be there.

Ben’s Bites is brought to you by Gauntlet

Most RAG systems fail in production. Gauntlet Night School on Wednesday, April 22 covers how to build one that doesn't — live and free - Register. Or go deeper: become an AI-native engineer at no cost. Cohort starts April 27. Apply now

Headlines

• Opus 4.7 is out - much better at vision (interpreting images) and efficient at using reasoning tokens. A new xhigh level of thinking now sits between “high” and “max”. I’ve been using it over the weekend at xhigh and didn’t face any issues (despite the general sentiment on Twitter that 4.7 is a regression).

• Claude also has a Design tab now - a canvas-like interface with chat on the sidebar to explore wireframes or create high fidelity prototypes. It asks you 5-10 questions via an interactive form and then gets building. I found the image → design workflow to be really good in the prototype mode. Has separate limits while in research preview, but expect the weekly limits to only last for 2-3 big generations max (on the $20 plan). Check out Peter’s demo across multiple use cases.

• Codex got a few updates: 1. Computer Use - it can now use apps on your Mac. I’m not completely sold on it and have some questions, but it does seem to work a lot faster than previous iterations of computer/browser use demos. Also, it works in the background, so your Mac is free for you to use. 2. Chronicle, an opt-in preview that uses recent screen context to build memories. 3. A bunch of plugins, including image generation, so that Codex can be the superapp to use AI.

• Factory AI is now valued at $1.5B after their latest raise of $150M. Try the new desktop app for Droid, Factory’s coding agent, with 50% off on Opus 4.7 till 30th April.

Subscribe now

My feed

• Your tools are full of data, but you’re still guessing. Backstory - Focus on what drives revenue, based on how your business actually sells.*

• Google AI Pro and Ultra subs now work with AI Studio - higher rate limits for vibe coding in the playground.

• Julius can now generate slide decks with charts and tables, exportable as pptx.

• Galaxy Brain - an operating system powered by local files. I’m an investor.

• Kimi 2.6 Code - a Claude Code-style terminal built specifically for Kimi K2.6.

• Moondream Lens - fine-tune a vision model to production accuracy in hours with as few as 20 images.

• Zapier’s AutomationBench measures real work done by models - CRM updates, inbox follow-ups, multi-step tool chains. No model has cracked 10%.

• How AI made me a builder without losing my taste as a designer.

• Quiver upgrades its models for vector generation - Arrow 1.1 and Arrow 1.1 Max.

• acceptmarkdown.com - checks whether your site returns Markdown correctly for agents.

• The 5 stages of AI grief.

• HeyGen open-sourced HyperFrames (HTML to MP4).

• Vercel was breached via an employee’s account on another AI product. Vercel says affected customers have already been contacted. Quick check to make sure you’re safe.

• Three OpenAI leaders left the company - Kevin Weil (CPO, then OpenAI for Science lead), Bill Peebles (Sora co-creator), and Srinivas Narayanan (CTO for B2B Applications).

• Skill drops: skill-creator, make-interfaces-feel-better, and gpt-taste for Codex.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Boarding my flight to SF very shortly, and I got an email to let me know - no WiFi today. Uh oh. I was kinda hoping my 11 hours uninterrupted hours without the kids would be productive for once (I’m usually a very OOO long-hauler, no internet). But I still have some work to polish this talk I’m giving on Tuesday.

I’m also in town looking to deploy $100k cheques to dev tools and infra founders, plus see some of my wonderful LPs and meeting new ones. Ben’s Bites Fund II has already started investing.

So my flight… I’ve had to hurriedly download a few local models so I can use my agents offline and I think, so far, Gemma 4: 26b is going to be my choice.

We’re so spoiled today with fast intelligence at our fingertips and it’s funny how used to the new intelligence levels we get

Local models are slow to boot up (you’ve got to be more mindful of what context is being loaded on startup (so I’m running with no-skills to get it to go faster, I can call the skills when I want — maybe I’d actually prefer to do that generally 🤔). And they feel pretty slow to do work, but only because of said spoils.

I’ve been in the weeds of context management recently because of the course I’m working on. And it’s been useful to just remind myself about how prickly it can be;

• If an agent runs web searches - presumably you didn’t read them, its gobbling up context from content you do not know is 1. right, 2. not ai-slop, and 3. by a source you’d recommend.

• Little (or big) lines of slop, misdirection, misinformation slip in to the context and compound over time

• Reaching ~60% of a context window is probably the limit of where you want to be

• Use other sessions as context-gathering sessions, if there’s lots of documents then create one summary file with the information (and try to read or at least skim it! - I am trying, promise)

• I don’t trust 1M context windows, there’s a great post by Thariq from Anthropic below about this window. I shouldn’t need my context for my tasks to need perfect recall beyond ~150k tokens, that’s a lot of words. Only until 1M context windows are the norm, the models dont forget anything and help clean polluted context along the way!

Anyway, got to head to the gate! This was a little different of an intro, let me know if you liked it. I need to share more as I’m learning (or diving deeper).

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there's more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• Claude Code’s desktop got a redesign. Brings many CLI-only features and more (like split windows for multiple sessions) to the desktop app. Big improvement, but still a lot is missing. It picks up some CLI sessions but not all, opening/editing files isn’t obvious, and it keeps asking for permission even with “bypass” settings on.

  • Gemini also has a native Mac app now. But it’s light on features - no Gems, no notebooks - and the design feels rough to say the least.

• New models - GPT-5.4-Cyber from OpenAI, fine-tuned for cybersecurity, with limited access to trusted partners. And Gemini 3.1 Flash TTS from Google - better voices, audio tags for controlling tone and pacing, and 70 languages.

• Routines in Claude Code are now in research preview - set up a prompt, a repo, and your connectors once, then run it on a schedule (or via API/GitHub trigger). Runs on Anthropic’s infra, so you don’t need your laptop open. Basically, extended cron jobs. OpenClaw calls these heartbeats.

• With the latest update to OpenAI’s Agents SDK, you can run Codex-style agents in production without building the whole harness yourself. You get sandboxed execution, computer-use, skills, memory, and compaction built in.

• Most RAG systems return wrong answers with complete confidence. Gauntlet's free Night School covers how production AI engineers actually fix that — setup, evaluation, the full loop. Wednesday, April 22. Register free*

Subscribe now

My feed

• Skills in Chrome let you save prompts as reusable one-click workflows that run on whatever page you’re viewing.

• Cursor can now respond with interactive canvases - dashboards and custom interfaces instead of just text.

• Resend shipped a new email editor with BYOA (bring your own agent). There’s a built-in LLM, but you can also MCP into the editor with your own setup.

• Sparkle v4 from Every - let AI organise your filesystem like you would.

• Daniel pointed an agent at 5 years of home-building emails (511 events, 690 documents, 170 finance records) and got back a full project timeline in ~$500 of Opus tokens.

• Impeccable v2 - the design skill for coding agents. v2 adds a CLI scanner (works without an LLM), a Chrome extension, and a /shape command that runs a design interview before writing any code.

• Using Claude Code - guide on session management, compaction, and the 1M context window.

• 30 min tutorial on building software with agents in Cursor.

• Lindy AI’s founder says GLM 5.1 will likely become their default over closed-source models for most use cases, saving them a bunch on inference (their biggest cost, more than payroll).

• OpenRouter now offers video generation models with one universal API across all video models.

• Copilot in Word now tracks changes and leaves comments.

• Windsurf 2.0 - Manage all your agents from one place and delegate work to the cloud with Devin.

• Gradient Bang - a fun multiplayer game with subagents in space. Built with Pipecat, Supabase, and open-source.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ll be heading to SF on Thursday — speaking at Stanford the week after. I’ll be delivering live the ‘how to use agents if you’re not technical’ course I’ve been working on (ahem, still working on).

I’ll be in town chatting to LPs (my 2 prev funds; 2.3x MOIC, 43% IRR (2023), and 4x MOIC, 28% IRR (2020)) and founders (I invest in AI dev tools and infra!) - let me know who I should try and meet or any cool things going on.

The latest leak from Claude-land shows Anthropic is adding lovable-like features for building full-stack apps right inside Claude. Codex apparently had a similar leak (since deleted).

There’s a new term going around for products that agents can use - headless SaaS. Box’s CEO says that “Enterprises will kick out vendors that don’t make it easy/economical for agents to use their product”. And you can’t just wrap your APIs into MCP/CLIs and call it a day.

Talks at AI Engineer were across the spectrum from “code is a liability” to “slow the fuck down”. Alex has a good recap.

Ben’s Bites is brought to you by Cloudera

AI is booming, but ROI is lagging. The problem isn’t AI—it’s data readiness. Most enterprises aren’t as prepared as they think. See where things really stand in Cloudera’s Data Readiness Index.

Headlines

• Claude Cowork is now generally available - out of research preview after 12 weeks with millions of users. Also, Claude for Word is in beta - draft, edit, and revise documents from the Word sidebar with edits showing up as tracked changes. (Team and Enterprise plans)

• New Claude Code features - /ultraplan lets you build and edit a plan on the web, then run it in your terminal. The Monitor tool lets Claude watch for events in the background instead of constantly checking - saves a lot of tokens. Plus, an advisor strategy on the Claude dev platform that pairs Opus with Sonnet for better performance at similar/cheaper costs.

• OpenAI added a new $100 plan. Their own wording is too complex, so I’m stealing Theo’s. Basically, $100 plan = 5x compute, $200 = 10x compute (where $20 plan = 1x). The bigger plans have 2x boost till May 31st, so effectively, $100 = 10x and $200 = 20x compute.

• LlamaParse was built to tackle your most complex PDFs with the highest accuracy & the lowest cost. Now we want to put that to the test. Submit your ugliest PDF (dense contracts, financial tables, etc.), and judge the output vs your current OCR to win a Mac Mini. Can your doc outsmart LlamaParse?*

Subscribe now

My feed

• Anyone can make a game now. Yoroll turns a rough idea into a branching, playable video-native game. Join waitlist to try.*

• Cursor cloud agents can now attach screenshots and demos to PRs they open. Your team reviews the artifacts directly in GitHub.

• Shopify AI Toolkit - manage your Shopify store with your favourite agent. Works with Claude Code, Codex, Cursor, VS Code, and more. (docs)

• Open Agents by Vercel - open-source template for building coding agents.

• How Missions work - from the Factory team.

• Agentic parenting and TradClaw.

• The anatomy of an agent harness and a tutorial/walkthrough to build your own harness.

• How I became technical AF.

• Personal wiki tools inspired by Karpathy’s knowledge base idea - LLMwiki, Wiki OS, Hatch, and GBrain.

• Evo, another Karpathy-inspired CC plugin, optimises your code through experiments.

• Ramp says their entire company is AI-pilled. They even built a new internal product called Glass to give every employee an AI coworker.

• getdesign.md - browse design systems of popular applications as markdown files. Preview instantly, install with one command.

• Aqua Voice - premium voice keyboard for every app on your phone. Now on iOS.

• Gitinspect - replace “hub” with “inspect” on any GitHub URL to chat with the repo. Runs in the browser, everything stays local.

• Cloudflare Sandboxes are generally available - comes with a terminal, interpreter, live preview URLs & secure credentials. Sleeps when idle and wakes on demand.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail sourced from Twitter

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

A mis-timed blog last week leaked Anthropic’s next model - Claude Mythos. Well, it is real and has massive improvements on benchmarks over Opus 4.6:

• 53.4% → 77.8% on SWE-bench Pro

• 65.4% → 82% on Terminal-Bench 2.0

but we are not getting access to it anytime soon. Why? because it is really good at finding and exploiting software vulnerabilities. On Firefox exploit generation, Opus managed 2 working exploits out of hundreds of attempts. Mythos hit 181.

It found many-decades-old bugs in critical software projects like OpenBSD (27-year-old bug), FFmpeg (16-year-old bug) and more.

Instead of releasing it publicly, Anthropic is giving 12 companies access to a preview version of Mythos under “Project Glasswing” to find vulnerabilities in critical software. Anthropic is committing $100M in model usage credits and $4M in donations to open-source security orgs under this project.

Theo made a video on this, and I like his point: “Mythos is to Opus what Opus is to Sonnet.”

I tweeted a list of companies that Meta has acquired in the past year without anything to show for it, and soon after, Meta released details about their latest model - Muse Spark. At a glance, it sits somewhere between Sonnet 4.6 and Opus 4.6. Not usable yet: API access is coming, and there are promises about open-source too (rip llama).

Many people are dunking on Meta for its not-so-frontier model release after spending billions and a year of silence, but I think it’s a good step ahead. Plus, have you used Instagram search over the past couple of months? It’s gotten really good courtesy of AI.

As always, good recap from Ethan Mollick on the state of frontier models: Google, OpenAI and Anthropic lead, Meta joins the pack for now while xAI has fallen off, and the best Chinese models are still 7-9 months behind.

ps: Factory’s desktop app is now out of beta. It comes with a cloud computer, the ability to use other apps on your device, and, of course, the ability to run and manage multiple Droid sessions easily.

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there's more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• Claude Managed Agents - You can use Claude’s developer console to build and deploy agents and let anthropic handle the infra for it, vs building it yourself. For example, Notion is using managed agents to build a “delegate tasks to Claude” feature. (Anthropic’s engineering blog on building this).

• Cursor has a new design mode to annotate and target UI elements in the browser. Plus, run Cursor on any machine and control it from anywhere, including your phone.

• Gemini app finally has projects - they call it notebooks. Similar features as Claude/ChatGPT projects - move chats in/out of notebooks, notebook-specific files and memories, with the additional feature to sync these notebooks between the Gemini app and NotebookLM.

• Clicky is an ambient AI buddy on your Mac. It sees your screen, talks to you and points at things to guide you (demo). Farza built (and open-sourced) it as a learning tool, but people are using it for everything.

• Choosing an accurate speech-to-text model is harder than it looks. Benchmarking one is even harder. See why standard word error rate falls short, and what better STT evaluation actually looks like.*

Subscribe now

My feed

• Chronicle: Cursor for slides. Never build a deck from scratch again. Turn ideas into stunning presentations in minutes.*

• OpenRouter Spawn - Deploy OpenClaw and other agents to the cloud of your choice. Works with all models on OpenRouter.

• Zapier’s SDK is now open to everyone. Programmatic access to all of Zapier’s capabilities. Free to use in beta. (docs)

• Kiro.dev (spec-driven IDE from Amazon) is bringing its startup credits program back for startups with up to 30 people.

• Cogito - Markdown editor for Mac. I’ve been using Clearly (recently updated) for the last few weeks to simply view and edit md files.

• Graphify - Turn any codebase or folder into a queryable knowledge graph.

• Pi and Mario (the maker of Pi) are joining Earendil, the company by the creator of Flask. The core harness stays open-source. New features will be a mix of enterprise & fair source (proprietary now, open-source later).

• Impeccable - Free design skills for coding agents with 21 commands to audit and fix common mistakes.

• Superset and Builder 2.0 - two new UIs for running parallel agents. Superset is more like Codex (terminal-first, worktrees), Builder is more kanban-style with Slack/Jira integration.

• CSS Studio by Motion - Make design changes by hand on your website in the browser, then pass them over to your agent for implementation.

• S3 Files from AWS allows storing data as a file system, making it easier for agents to use.

• Every is running two parallel org charts - one for humans and another for each employee’s openclaw agents.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

But OpenClaw founder, Peter, is a (good-kind of) maniac who’s trying to get gpt5.4 working like opus in the tool (since OpenAI acquired him).

Unfortunately it’s just left the whole community confused as to where and when they can use their Claude Code subscriptions outside of the cc harness itself. Myself included.

There’s a new take on knowledge bases that I think is very interesting by Andrej Karpathy. Memory and file organisation for retrieval got a big focus with the OpenClaw hype but this approach groups things topically and then enhances with summaries, backlinks and wikis.

Related, Farza built Farzapedia, a personal Wikipedia generated from 2,500 diary entries, notes and messages. Built for his agents to crawl whenever needed. Karpathy’s take on Farzapedia

OpenAI acquired TBPN -- good pod to get a peek into the heads of the hosts. A lot of people on X are trying to sound smart with ‘this is why they acquired them’ but I’m more in the Ben Thompson camp… I don’t really know why either party needs each other here? TBPN is loved, growing, and making a small fortune with their (also loved!) ad business. Why does OpenAI need them? Other than the fact that they don’t like traditional media and could use TBPN as their main channel to people that matter.

Ben’s Bites is brought to you by Reevo

The best sellers aren’t the ones with the most sophisticated tech stack. They’re the ones who spend the most time selling. Reevo is the AI-native revenue intelligence system and CRM that every team deserves. One Platform. From prospect to close in a single tab. Go Stackless. Reevo.ai

Headlines

• Google released Gemma 4. 4 new open-weights models - two for powerful laptops and desktops (26B MoE and 31B dense), and two tiny ones for mobile phones (2B & 4B). Relevant for companies that want to fine-tune them, not for direct use.

• Cursor-3 adds a new standalone window for running agents. Has some neat additions for local to cloud handoff, working across different projects.

• Anthropic’s run-rate revenue hit $30B, up from $9B at the end of 2025. Lenny’s podcast with Anthropic’s head of growth is worth a listen — they added $6B ARR in February alone.

• Adapt is your company’s AI computer. It connects to your tools and team, autonomously learns your business, and gets smarter with every task. Pull live data, run scheduled workflows, and deploy internal apps everyone can use, all from Slack.*

Subscribe now

My feed

• Salmex I/O — Won't forget you. Gets things done. Answers to no one. And is actually safe! What a local AI agent should be.*

• What could “devrel” mean in a world with slot machines spitting out code and opinions?

• agent-browser by Vercel can now also work as a browser agent with a chat-bar on the side.

• Turning any documentation website into a filesystem for agents to navigate with tree, grep, cat, etc.

• The CEO of Brex claims he is running the $5B+ company with OpenClaw.

• Growth ideas that helped Michelle scale Warp to 500k+ weekly active developers.

• The era of web apps is over - Bret Taylor

• Deep Extract adds an agent in the loop for extracting data out of complex documents that verifies and corrects itself.

• Samuel is building skills (like this AI SDR one) for the Replit Agent and sharing them on X. Cool way to market agents to non-technical users.

• Simon is the most up-to-date person on LLMs. He was on Lenny’s pod to talk about the changes in AI coding over the last few months.

• Intercom is making its agent Fin available via API for companies with contracts starting at $250k per year. (read more)

• Loophole translates your natural language moral beliefs into codified laws and then tries to break them.

• Ex-OpenAI board member (left after the Sam Altman firing fiasco) argues for retiring the term AGI in favour of specific milestones.

• You can now use your agent to apply for a job at Every.

• Autoagent - An agent to help optimise agents.

• Planning is dead

• A YC partner one-shotted a tool to manage his inbox with voice. Another YC company launched an AI secretary that manages your email like you.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Re unreleased/hidden features (because I know you wanna know); “Proactive” mode (AI works autonomously without prompts), frustration/anger detection via keyword patterns is marked ‘negative’ in their analytics, a Tamagotchi-like “/buddy” companion, background daemons (like OpenClaw uses), undercover commit-hiding mode, and more feature flags/roadmap items.

The community ran rampant cloning the codebase on GitHub (which Anthropic since sent DMCA notices to get them taken down). But some developers ported the code to other languages (Python and then Rust) - which has copyright greyness and if Anthropic push to get them removed, may bring up questions on their own copyright issues.

Since it was April Fools yesterday, many launches are indistinguishable from real ones - I think it’s getting harder to spot in the AI age too.

But some folks use it as an opportunity to launch things, like Gumroad replacing their CEO with an AI Agent. Sahil (the founder) is a friend and previously invested in my last company. He’s super smart and I don’t think this is a prank at all. He’s the kind of person to try wacky, out-there things to see if they stick.

Some companies actually do April Fools well, like ElevenLabs. Last year, they made dogs talk, and this year, they are partnering with the 3000-year-old ElvenLabs.

Ben’s Bites is brought to you by LlamaIndex

Your agent can write code but it can’t accurately read PDFs. Classic.

LlamaParse 🦙 reads docs with 99%+ accuracy on complex PDFs–like SEC reports, invoices, or research papers. Give your agents the context they need.

Sign up today to get 20,000 free credits with coupon: BENSBITES20 (ends 4/10)

I invested in LlamaIndex when scouting for a16z a few years ago

Headlines

• Claude Code now renders a full UI in your terminal to solve the flickering. It shows “N new messages” when you scroll up, gives you full mouse support, and constant memory use. It’s experimental for now; upgrade to the latest version and run using “CLAUDE_CODE_NO_FLICKER=1 claude”.

• OpenAI closed its $122B raise at a $852B post-money valuation. It’s making about $2B a month in revenue, 40% of which comes from the enterprise. The blog post outlines a lot of their plans that were previously just rumours/leaks from various newsrooms.

• Slack is turning Slackbot into a desktop companion. Select anything on your screen, pass it to Slackbot, and it can answer based on the context from your Slack workspace. It has 30+ new capabilities, including meeting transcription, reusable AI skills, MCP client and a native CRM.

• Softr AI builds the tool your business needs, not a prototype. It wires the whole thing: database, logic, logins, permissions, & security (even hosting) with a visual editor to fine-tune. Think client portals, internal tools, CRMs, and more. No code required. 👉 Start today with 200 free AI credits.*

Subscribe now

My feed

• Attio - The AI CRM that does the work for you.*

• AgentOS by Rivet.dev - open-source operating system for agents. ~6 ms coldstarts, 32x cheaper than sandboxes.

• Mario, the founder of Pi, asked around for something like Google Docs, but for markdown. He then vibed Jot in an hour.

• Your agents can now access Supabase’s full documentation as a virtual filesystem. I think Mintlify could offer this to their customers, similar to how they added the “Copy as Markdown” button to the docs hosted via Mintlify.

• Vibe coding SwiftUI apps is a lot of fun.

• Simple Mac app to run skills on a schedule via Claude Code.

• Exa’s new product, Monitors, returns fresh results from the web for your query on schedule. Feels relevant for building agents.

  • Yutori’s Scouts do something similar for humans. They also launched Yutori’s desktop app for local-only access to logged-in websites when creating these “what’s new for your query” reports.

• The case against personal software.

• Cloudflare launched a successor to WordPress with a focus on plugin security.

• How Zapier measures the AI fluency of every hire across all departments.

• Block wants to move away from traditional human hierarchies and transition to one guided by an intelligence layer.

• Microsoft is working towards adding OpenClaw-style personal agents to M365.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Railway accidentally let unauthenticated users access data that should’ve been behind an auth wall.

Mercor AI has allegedly been breached.

Claude Code’s source code has been leaked - and the community are going crazy saving copies on GitHub. Docs from the codebase.

And, Axios, with 100M weekly installs got compromised through package manager, npm when one of the lead maintainers’ GitHub accounts was hijacked. npm has removed the malicious versions now. (Claude code uses axios)

Like manufacturing, code has a supply chain.

When working with software, it relies on other software. Instead of writing all the other code into your project, you install a package through a package manager - agents do this very frequently on your behalf.

One package, Axios, was compromised, which means if an agent (or you) ran the install command, a malicious package is now on your computer.

This will stress the importance of sandboxes. Tools like Claude Cowork and Codex do this for you by running commands in a sandbox, a computer with a copy of your current folder isolated from your computer. So if any bad code sneaks in, it doesn’t mess up your actual stuff!

I sent this to my agents this morning:

there’s been a security breach https://markdown.new/socket.dev/blog/axios-npm-package-compromised

make sure this computer and my mac-mini have not been compromised

What am I building this week?

• I’m purposefully trying not to build too much (hence no Ben’s Builds email last Saturday) because I’m focusing on this course. It’s taking shape now, and I hope to send out some preview lessons asap. I’ll be presenting a version of this to Stanford students in SF next month.

• I really do want to finally spin up my own email client, probably by cloning this, made by a YC partner.

• Building security or sandbox-related developer tools or infra? I invest 👋

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there’s more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• Computer use is now in Claude Code. Claude can interact with your computer using the UI (like we do) to test apps or do tasks. Available in research preview on Pro and Max plans—expect it to be slow, clunky and expensive. Separately, Claude Code auto-fix works in the cloud, via web and mobile sessions. It watches PRs, fixes CI failures and addresses comments remotely.

• Projects.dev by Stripe lets agents use third-party services from the CLI. Run a command, and it creates an account, gets an API key, and sets up billing with partnered apps like Posthog, Supabase, Clerk, PlanetScale and more. Developer preview is live, open to everyone soon. I got access, and it’s pretty great, much simpler than using multiple tools and connecting them.

• Gemini Live is powered by a new model now - Gemini 3.1 Flash Live. Takes in anything—text, images, audio and video to output text & audio natively. Better than GPT-Realtime 1.5 and others on following complex instructions given via voice. Available for developers too. Gemini now also supports importing your entire chat history from other AI chatbots (with no way to export your Gemini chats at all). Diabolical.

• Codex has plugins now, i.e. a bundle of skills, app integrations, and MCP servers for building reusable workflows. They also created a plugin for Claude Code that lets you use Codex inside CC (how to use it).

Subscribe now

My feed

• Remodex lets you control Codex (running on your Mac) from your iPhone. Pico lets you do the same for pi-coding-agent running on any machine, via any mobile.

• Shopify released a suite of free tools to create images in a new mobile app called Tinker. It lets you create images and videos like social media posts, product staging, virtual try-ons and more.

• Litter - Codex on your phone.

• here.now sites can now connect to external services: Supabase, OpenRouter, Stripe, and Resend. No backend needed. One of my favourite tools (I’m an investor) just got even better! I’ll cover building with this soon.

• Plus One by Every - A hosted OpenClaw that lives in your Slack, pre-loaded with skills, workflows, and connected to other Every tools like Cora (email), Spiral (writing), and Proof (docs).

• Everyone is building a software factory; no one has figured it out fully yet.

• Hyperbox - rent Mac-mini’s as your virtual computer (re the security topic today!)

• Vercel published excerpts from an internal talk on how to agent responsibly.

• OpenClaw will soon be an MCP

• How Claude Cowork’s design lead uses it to collect and summarise user feedback to decide what gets built next.

• Chroma and Intercom have both trained custom models for their use cases. Chroma’s Context-1 is a better search agent, and Intercom’s Apex 1.0 helps their agent Fin achieve a higher resolution rate. Intercom’s CEO makes the case for vertical models.

• Cohere Transcribe - 2B text speech to text model with faster and better performance than most similar-sized open source models.

• Warp stopped buying SaaS and moved everything to agents, skills and just-in-time apps. Saving $10k+/year on cancelled subscriptions.

• Users who talked with Macy’s new AI Chatbot spend about 4.5x more than users who don’t.

• Daniel thinks online courses should position themselves as training for agents now.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

The most common tools are in the form of CLI. Agents communicate in text, CLIs are text in/text out, so it’s a natural fit. A CLI is a text-based way to control software. You type a command, something happens.

Here’s a simple example - organising files, using the bash tool.

"Rename all 400 product photos to match our SKU format, resize them to 1200x1200, and sort them into folders by category."

First the agent lists files to understand what it’s working with.

COMMAND:
ls ./product-photos/

OUTPUT:
file_1.jpg
file_2.jpg
file_3.jpg ... (400 files)

• ’ls’ is the command for ‘list’

• ./product-photos/ is where the files are located

Then creates the folder structure

COMMAND:
mkdir -p ./output/{shoes,bags,jackets,hats}

OUPUT:
output/
├── shoes/
├── bags/
├── jackets/
└── hats/

• ‘mkdir’ is the command for ‘make directory’ (directory is a folder), here it’s creating 5 - output, output/shoes, output/bags, output/jackets, output/hats

• flags modify what a command does: -p here means ‘create any missing parent folders too.’ So if ./output/ doesn’t exist yet, it’ll make that too

Then resizes the images

COMMAND:
mogrify -resize 1200x1200 ./product-photos/*.jpg

OUTPUT:
400 images resized ✓

• mogrify is an image-editing tool that edits files in place (overwrites the originals)

• *.jpg is a wildcard pattern meaning “all files ending in .jpg”.

Then renames and sorts each file

COMMAND:
mv IMG_0291.jpg ./output/shoes/SKU-1042-BLK.jpg
mv IMG_0292.jpg ./output/bags/SKU-2187-TAN.jpg
mv IMG_0293.jpg ./output/jackets/SKU-3301-NVY.jpg
... repeats for all 400 files

OUTPUT:
400 files renamed and sorted ✓

• 'mv' is the command for 'move' (which also renames the file when you move it to a new name)

Then it verifies the result before sending back to you

COMMAND:
ls -R ./output/ | head -20

output/shoes/
SKU-1042-BLK.jpg
SKU-1043-WHT.jpg
SKU-1044-RED.jpg
...(112 files)
output/bags/
SKU-2187-TAN.jpg
SKU-2188-BLK.jpg
...(89 files)

• On ‘ls -R ./output/ | head -20’. The | sends the output of one command into another, ‘head -20’ just means 'show me the first 20 lines.

It does all this in seconds. It would take you a couple of hours manually.

This is one CLI, called bash, the general-purpose command line that comes with your computer. But there are purpose-built CLIs for specific jobs too:

• Stripe CLI — pull revenue data, manage subscriptions, test payments

• Playwright — control a web browser: navigate, click, fill forms, take screenshots

• AWS CLI — spin up servers, manage databases, scale infrastructure

• Vercel CLI — deploy a website live in one command

Each of these is a separate tool an agent can use. The file organising example used one tool (bash). But give an agent the Stripe CLI too and now it can pull your revenue numbers. Add Playwright and it can browse the web. Add Vercel and it can deploy what it builds.

That’s what “tool use” means. The more CLIs you give an agent access to, the more it can do. Your job is to make sure it has the right ones for the task.

It all sounds a bit technical, and it is, but you’d only see those raw commands if you’re using a terminal or watching them fly by in tools like Claude Code. They’re present even when you don’t see them.

If an agent like Cowork is doing a task, you can click to expand what it ran and see the detail — like this example listing files to find recent fund updates.

Every agent is running commands like this under the hood. The interface just hides and abstracts them away.

Headlines

• Claude Code launched auto mode, a middle ground between manually approving every action and skipping all permissions dangerously (how they designed it). Claude connectors for work tools are now available on mobile too. They are also cooking something called auto-dream for compacting memory overnight. Claude Code can now use iMessage to text you and others. (see docs)

• Sora is shutting down. OpenAI is killing its standalone video generation app along with the API. Its $1B deal with Disney is also cancelled as a result. The Information reports that OpenAI is culling its side projects and focusing on a few key bets, with a new model codenamed Spud.

• ARC-AGI-3 launched with 135 mini games, nearly 1K levels, all human-solvable. But all models, when given basic prompts, score less than 1%. They have 25 games publicly available to play (as humans) and don’t tell anyone that I spent 4 hours on them yesterday.

• Google released the Pro version of Lyria 3, extending the music generation from 30 seconds to 3 minutes. It’s available in both the Gemini App and AI Studio for developers.

• The Figma canvas is now open to agents. You can now use AI agents to design directly on the canvas using the new use_figma MCP tool.

• Why Portkey is making its latest Gateway launch completely open source.*

Subscribe now

My feed

• Chronicle – Cursor for slides. Turn ideas and notes into stunning, professional decks in minutes.*

• Paper Snapshot - Snapshot your live website and paste it into Paper as editable HTML/CSS layers.

• Ghostwriter by Sierra - Chat with an agent to build more agents.

• Mario, founder of the popular open source agent Pi, wrote a post yesterday, “Thoughts on slowing the fuck down“, that says software quality appears to be declining as more companies rely on agents.

• Building CLIs for agents - Eric from Cursor wrote a thread on making CLIs that actually work for agents. ElevenLabs has already made their CLI agent-friendly using these tips.

• Building deep research that works from your CLI with BrowserBase. (resulting code)

• Hark – New AI lab from Brett Adcock (yes, the Figure robotics guy). 8 months in stealth, focused on "the most advanced personal intelligence" paired with next-gen hardware.

• GitHub has been going down wayyy too often these days. Plans to fix it and alternatives are starting to show up.

• How USV built a team of internal agents that live in their group email threads and learn from team feedback.

• Feynman - Read papers, research and get cited meta-analysis for your question from your CLI.

• Brave registered the .agent TLD and is making it a community effort. I tried to reserve 10 domains 😬

• Lil Agents – Tiny AI companions that live above your dock. Each one has its own Claude session and mini window. Now open source. Adorable.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

But how do you actually give it instruction files to write/design/code/whatever like YOU want it to?

A pattern I see a lot is getting your agent to interview you.

Your task is to interview me and get all the information you need to [your task].

Then add a specific ask based on your goals…

Come up with a number of interview questions for me about landing page designs I like, why I like them and what I don’t like.

Ask me a number of questions on how I think about sales copy, ask me for some examples I like and dislike and why.

Interview me about writers I like, which books/posts I love from that author and why.

I recently did this with the course I’m working on.

I hate courses. I don’t think the majority of them do all that much teaching. They walk you through steps to mimic getting you to the end goal. But once you ‘graduate’ (not many people do), you’re on your own.

But real life is never that straight forward, you’ll always hit bumps in the road and courses don’t give you the knowledge to navigate them.

I get stuck with blank page syndrome. I need something, anything, to start me off - even if it’s AI slop.

So I asked my agent

I am writing a course on 'Becoming a builder'. A helpful guide for non-coders to learn how to work with and steer agents, understand systems of code projects (not specifics of how to code) and learn whilst building.

I want you to interview me so we can flesh out the course content map - a hierarchical overview of the topics to cover, with some bullets on the points to cover.

Ask me one question at a time, I may disregard questions but I will say why I don't think its relevant. 

Feel free to probe further if you dont have enough context for any section.

Really try to understand what and how I'm teaching based on the end outcomes I've specified.

The agent asked me 20 questions and I spoke my rambling thoughts back. I was genuinely surprised how often it would remember to probe me, or ask clarifications like ‘do you actually mean to go down this route or this one’.

All in all a very helpful exercise in getting off the blank page. I now have a number of sections with ‘what this covers’. Even if on first glance I know I’m going to remove/merge/edit a lot, I’m moving forward.

What am I building this week?

• I’m furiously working away on the course I mentioned above, named Fork Off.

• I want to revisit my OpenClaw/personal agent memory system - has anyone found one that they absolutely love and swear by?

• I really want to make a YouTube wrapper for my kids where I can pre-approve channels I let them watch. Fuck CocoMelon, ASMR, cutting coloured sand and all that crap. YT’s algo just constantly surfaces these. Also if you make a YT video for kids, please put the thumbnail scene at the start of the video - or you get meltdowns 🙃

Ben’s Bites is brought to you by Reevo

Go stackless and get back to selling. Remember when selling meant talking to people? Before the tab-switching and endless sync errors. Reevo brings it all back to one platform. Prospecting, calls, pipeline, and reporting all in a single tab. From prospect to close. Go Stackless. reevo.ai*

Headlines

• You can schedule recurring cloud-based tasks on Claude Code, and you can now enable Claude to use your computer to complete tasks. It uses your connectors first, but if there's no connector, it’ll use your computer to open the app (but your computer must be on!) Plus projects are now available in Cowork.

• Long-running agents designed to automate large software tasks like building applications from scratch with Factory Missions. This is genuinely the closest feeling of AGI I’ve ever had. You spend decent time planning your mission but then it just does everything end to end.

• ChatGPT now has a library of the files you upload, making it easier to reference them. OpenAI is also planning to simplify its product experience and launch one “superapp” - much like Claude has done with their Desktop product.

• Cursor launched Composer 2 as their latest ‘in-house’ coding model. It came to light that the model was a tuned version of Kimi’s 2.5 open-source model (which they failed to mention, which caused some rumblings on X.) They boasted about their high scores on their own benchmark, CursorBench - but only compared their scores against Claude Code/Codex (not any other harnesses which outperform them), which feels weird considering they are a harness themselves. They also released ‘Glass’, which is their new interface that follows the 3-column layout that lots of apps are using.

• SpaceX, Tesla and XAI launched TERAFAB, the largest chip manufacturing facility ever (1TW/year). This post from Sequoia Partner Shaun Maguire puts forward the idea that everyone is sleeping on XAI, and how it will win in AI.

• New model, worse benchmark. Plot twist: the truth files were wrong. AssemblyAI found their AI penalized for transcribing things correctly that human labelers missed. Live workshop March 31 on why WER breaks and how to fix your eval pipeline.*

Subscribe now

My feed

• Speechmatics – STT for voice agents. <300ms latency, high accuracy at conversational speed, 55+ languages. BB readers get $200 free credits.*

• Codebase to Course - a skill to make learning codebases more visual and interactive.

• Luma Labs has a new image generation model: Uni-1. Think Nano Banana with a canvas and multiple outputs per turn. It’s good but slow because of the many outputs (sometimes 10+ for a single message).

• GPT 5.4 is awful at frontend design - but this guide apparently makes it better. And they’ve added a frontend-skill to use in Codex. I’ve not yet tested it - mixed reviews on X.

• ~99% SOTA memory system for agents.

• ‘The Minimalist Entrepreneur’ is a great book for founders, and now Sahil’s turned it into a set of agent skills.

• A podcast with Simon Willison: Engineering practices that make coding agents work.

• Must read/listen: The End of Coding: Andrej Karpathy on agents, autoresearch, and the loopy era of AI.

• How to use Claude Code to design, by one of the Tailwind founders.

• Ghost Pepper - 100% local hold-to-talk speech-to-text for macOS.

• Deploy a team of OpenClaw agents securely.

• Lovable is kinda pivoting from an app-making tool to a general agent.

• AI style guides: How to help AI write like you

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

This week was pretty unproductive from a building POV as I went back to my hometown, Cardiff, to speak at my old school about business and AI.

To my horror surprise, almost all 16-18 yr olds I spoke to only use ChatGPT and only for school work. I was half expecting some kids to have used it to build something. We have more work to do!

I’ve been muddling around for a while now, at least since our third, Poppy, was born in December. It’s time to get some ‘proper’ work done now.

I’m leaning back into educating non-coders how to build with AI. Shock.

Subscribe now

What did I build this week?

1. Fork Off

Name tbd but I like it 😂.

This will be a course I’ll open up once a quarter to teach non-coders how to build things with AI; apps, automations, agents (oh my!). Why once a quarter? Because everything moves so fast. I can tweak the content without worrying about new folks joining and getting stale content.

It’ll take a few weeks to get the content right but I spun up a custom course platform for this using Droid. You’ll be able to log in, see progress, the usual stuff.

I procrastinate on courses a lot because I don’t think step-by-step tutorials are useful and I need to be able to visualise how it looks, feels and most importantly, does it teach what I am really trying to cover (ie teaching you how to build so you can feel confident outside the course).

Why ‘Fork Off’ - a play on forking projects to remix them for yourself. I’m a simple man.

2. X Bookmarks Search

I now rely on X bookmarks for nearly all the content I put in our regular emails, plus a bunch of stuff I want to try/use/copy later but it’s impossible to search easily, filter etc.

I just spun up a quick web app that does exactly that. I may make it public. Need to work on the filters!

3. Email Triage Agent

I have one already. I set it up once to label/archive emails for me but I need to dramatically improve it. So I thought this was a good chance to try out Replit Agent 4.

It was 💩.

I can only hope it was my own skill issue but so many little things just did not work. I had a little running note of all the issues I was running into;

Then I gave up. I’d have finished if I used my usual Droid/Pi tools.

I’ll be back to test this again and do a proper breakdown. And finally get an email agent working?! 🤞

Builds from last week…

• Cookbook site I haven’t touched since last week - oops

• Become a builder - I published the first cookbook last week and will be folding this into the course

https://gists.sh/bentossell/a4e5e7048e8a355ec56cf3db86169ae2 

What’s in my stack - tools, skills, instructions, models

New tools I’m using:

• Chops - a light mac app to manage all your skills

• Claude Cowork - I ran a poll in the newsletter this week and most of you want me to do a breakdown of Claude Cowork. I’m now using it daily, it’s not my go-to tool (and obviously limited to just claude models) but I think there’s a lot that can be done in this tool. Plus 2 leaks; folders and text it!

Other tools

I hate AI writing, and consider myself a poor writer. I don’t use AI writing much except to get me off the blank landing page problem. I saw these recently which I’m starting to play around with.

• Every’s AI style guides - which comes with a prompt for your agent to interview you and nail a style guide…let’s see!

• Jack Butcher has a way with words + images. I pointed my agent at this repo to write the copy for forkoff - which I will change but it’s a nice starting point.

• Tropes is another tool/set of instructions to avoid AI slop writing too

I think some unique combo of this packaged as a skill is something I should try and figure out properly 😊 to be continued…

Chat with me

If you know a builder that’d find this useful, feel free to forward to them.

Have a great weekend!

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

Every model has a system prompt - instructions from the model/product makers. Your .md files get added to this. The model now has its system prompt + your instructions pre-loaded.

CLAUDE.md is specific to Claude products. AGENTS.md works across Codex/Droid/Pi/most other tools. You can ‘symlink’ these - ie link my AGENTS.md to CLAUDE.md so that I only ever deal with one set of instructions and any agent will pick it up correctly. (Just ask your agent to symlink them)

What goes into your AGENTS.md?

We previously thought including your tech stack, key files, etc., as like a mini-map for your agent was the right approach. That’s what agents add if they create it.

But there was a study that showed it hurt performance and increased cost by 20% (using extra tokens). The agent can figure out the tech stack, key files, commands, and architecture very easily and quickly.

Instead, it should be pretty empty. It should be your preferences and nudges to correct agent behaviour.

  - When building, open a browser with agent-browser skill and test before sending me a URL (to catch bugs)
  - Use the Exa web search tool for web search
  - Always write planning files in ~/[project-name]/plan/
  - I can't code, so explain things in simple terms
  - Record a video of your output so I can see exactly what you tested

I picked up a tip from Dex, who says wrapping sections in conditional blocks is helpful:

<important if="simple web page">
  - No spec needed
  - Create 3 designs before choosing one
  - Must have dark/light mode switcher
</important>

I often switch between spinning up simple sites and more complex apps. I noticed my agents keep writing specs and testing in the browser, which is unnecessary. So I’m adopting this.

You don’t need to mention skills you’ve installed, as their ‘frontmatter’ (the skills’ name and description) is also pre-loaded alongside your AGENTS.md.

If you’re using ChatGPT/Claude Desktop apps these instructions still work. Paste them in ‘personalisation’ / ‘preferences’ / or ‘instructions’ - in your settings.

AGENTS.md also get dynamically loaded as an agent navigates through folders.

my-project/
├── AGENTS.md              ← root instructions (always loaded)
└── docs/
    └── AGENTS.md          ← loaded when agent works in /docs

If it was helpful, reply and let me know 😊

What am I building this week?

• I’m currently in Cardiff doing talks with 16-18 yr olds on starting companies and AI. Making me think I should do proper workshops for these kids.

• I just killed my email triage bot. Going to try a different approach. Tried Replit Agent 4, but just did not work. I’ll record a video on this build + Replit review

• I’ve been using keep.md to bring all my saved items into one feed (that gets filtered for this newsletter) - and I’m going to make the feed public

• I need to cut my AGENTS.md down + add conditional blocks (as above)

• I may do a custom ‘claude for chrome’ extension…

Ben’s Bites is brought to you by Viktor

Your next hire isn’t on LinkedIn. Viktor is your last hire. He has his own computer, lives in your Slack, watches how your team works, and starts solving problems before anyone notices them. Writes code. Builds apps. Runs campaigns. No triggers. No workflows. It just works. Try Viktor free, you have $100 up front.*

Headlines

• Google launched the new Stitch. They’re calling it your ‘vibe design partner’. Gemini is great at generating UI, unlike other models (ahem, openai), and now paired with their own harness with a bunch of features; AI-native canvas, design agent, voice, instant prototypes, design systems and DESIGN.md (handily links with our intro! Take a look at it 😊 ). Here’s a good overview video of the tool.

• Claude launched ‘Dispatch’ for Cowork. Once you connect, you can send messages from your mobile to the Claude Desktop app (i.e. work with files on your desktop). It can also launch Claude Code sessions. I got this working, but I had to approve some permissions on my Mac first (“git” & “claude” would like to access your files). They also published a report: What 81k people want from AI

• GPT-5.4 mini is available in ChatGPT, Codex, and the API (5.4 nano available in the API only). Roughly Sonnet 4.6 intelligence, 3x faster and ~70% cheaper. OpenAI also launched a challenge to build the most efficient pretrained model under extreme constraints.

• Midjourney’s v8 model is in early testing. It’s 5x faster and better at rendering text, but nowhere near nano-banana. It wins in the aesthetics as always, though.

• Runway’s new real-time video generation model created instant HD videos on Nvidia’s Vera Rubin hardware with time-to-first-frame under 100ms.

Subscribe now

My feed

• AssemblyAI: Real-time STT that keeps up with human conversation. Multilingual with live diarization. Built for voice agents. Start free.*

• Unsloth Studio - Open-source web UI to train and run LLMs. Training using your Mac is coming soon.

• Lenny put his entire newsletter archive + podcast transcripts into markdown files (which agents love). People have built RPG games, a parenting wisdom site, infographics and more. We should do this for Ben’s Bites data!

• A sufficiently detailed spec is code - spoiler: no, it’s not! The argument is that writing a specification detailed enough for an AI to reliably generate code from it becomes as complex as writing the code itself. It also claims that AI-generated specs lack the thoughtfulness that made a specification valuable.

• Gamma launched Imagine (visuals, logos, etc), AI-Native templates and connectors.

• Browserbase agents get 1000 searches/mo for free with their new Search API powered by Exa AI.

• 1Password launched ‘Unified Access’ to help agents get access to credentials.

• 2M people read why we banned React’s useEffect

• Will AI kill spreadsheets

• The agentic coding flywheel

• OpenOats (fka OpenGranola) - A meeting note-taker that talks back

• Chops - a skill organiser app - I just downloaded!

• Humanoid Atlas shows where all the parts to make humanoid robots come from - pretty cool data visual

• Stripe launched the Machine Payments Protocol that lets AI agents programmatically pay for services. And Visa’s Crypto Lab launched a CLI

• A mini profile on one of my portfolio founders

• OpenAI has a new focus (on the IPO)

• How we run a code factory

• How to price your AI product

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

People in the developer/tech community are openly talking about leaving social events early to get back to their AI agents, skipping drinks to stay sharp, lying awake thinking about what they can run before they fall asleep, and constantly spinning up new projects and ideas. There’s a shared, unspoken anxiety driven by the relentless pace of AI progress, where every week makes last month’s workflow feel obsolete, and the window to be “first” at anything feels like it’s shrinking by the day.

It’s absurd.

I don’t want my emails to add to this. I want to spark ideas. A new tool you can use, a workflow you can copy and generally interesting posts from others.

We felt this when we reduced sending 5 emails a week to 2. But often we get carried away adding too many things here that just don’t matter.

We’ll be more conscious of that.

Subscribe now

Here’s an alternative version of today’s post - at the bottom of it you can vote on which you preferred. I’d really appreciate it!!

Headlines

Codex now has 2M+ weekly active users, and OpenAI API use is 20% up since GPT-5.4 was released. Fidji Simo revealed this in her post about launching OpenAI’s deployment arm for enterprises. I’ve been using Codex here and there - mostly testing. It’s a better user experience than all others at the moment, and if you use ChatGPT, Codex shouldn’t feel scary to use. Subagents are also now live in both the Codex app and the CLI

Manus (recently acquired by Meta) launched a desktop app, My Computer, to compete with Codex/Claude Code/OpenClaw etc. It was very fast (using the 1.6 lite model) but didn’t actually get the task right in my testing. I asked ChatGPT, Codex, and Claude Cowork the same thing; for all of these companies, fill in the PDF with their information, download the files. ChatGPT gave me a link to download the files, Codex did it and saved the files to my computer, as did Claude Cowork. Cowork was slowest. Manus was fastest but didn’t fill in the PDF correctly.

Jensen Huang says Nvidia expects to generate $1T+ in sales from its flagship AI chips through the end of 2027, after previously forecasting $500B by 2026’s end. They also released NemoClaw, an open source stack that adds privacy and security controls to OpenClaw. [repo]

Claude’s 1M context window is now generally available. I’ll hold off on my take until further testing - a few folks taking either side of its good/bad.

The best sellers aren’t the ones with the most sophisticated tech stack. They’re the ones who spend the most time selling. Reevo is the AI-native revenue intelligence system and CRM that every team deserves. One Platform. From prospect to close in a single tab. Go Stackless. Reevo.ai*

My feed

• Vibe coding, the term, is being phased out by ‘Agentic Engineering’ and Simon Willison created a guide on ‘what agentic engineering is’ - It’s still engineer-focused, should do one for less-technical folks?

• Travis Kalanick of Uber fame is back with a bang. He introduced (in a very long post) his new company, Atoms (which is actually 8 years old). It focuses on “digitizing the physical world” through robotics, sensors, automation, and physical AI. He went on TBPN for an interview

  • One of my favourite individuals and LPs, Om Malik, wrote a great piece “The Return of Travis Kalanick: Fact & Fluff!” because let’s be honest, who gets what he’s doing from ‘digitising the physical world’ tagline?!

  • He also wrote a great piece on the OpenClaw craze, too. Touching on the socio-cultural importance of the Claw movement: Lobster Boil

• Productive individuals don’t make productive firms

• Vibecoding is my passion by Ryan Hoover. He says how vibecoding is becoming a form of self-expression rather than just a means to an end (which I agree with). It’s a form of entertainment that (should be) enjoyable to take part it. We’re all software painters now.

  • Whats defensible in Ryan’s view? Social graphs, distribution, licensing, data, and hardware.

• Most AI tools do one thing. Viktor does your job. Not one part of it. The boring parts, the repetitive parts, the “I’ll get to it later” parts. Finance, ops, marketing, engineering. It lives in Slack, connects to 3,000+ tools, and delivers real work. Try Viktor free.*

Tools

• Facing the 2027 Opsgenie sunset? Join incident.io on 3/17 for dev-led migration hacks & scalable workflows that go beyond paging. Register for free.*

• If you wanted to have your own personal agent like OpenClaw but you’ve tried, or its too tricky - whatever the reason. The simplest packaged solution with the most power is Nebula. Built by a very successful founder, Furqan (from AppLovin). It just handles a bunch of the complexity for you - virtual computers, orchestrating other agents, integrations etc. And you can get started for free. I’d definitely give this a go.

• As personal agents like OpenClaw are seeding a new ‘buy an agent in a box’ model, we’re going to see a lot more vertical focused agents like this AI CMO one. Kind of like a ‘mission control’ for all your marketing; seo/geo/writer etc. I hate having to sign up for a new tool the second you’re dropped into it to see what it does.

  • The flow for these kinds of companies; make a focused skill for your agent to do a task, then turn it into its own agent.

• Another always-on agent with automations launched, Adaptive. I’m going to have to test all of these side by side to give a proper breakdown. They all feel very similar.

• Factory released Factory Analytics - tokens → usage → commits → pull requests → shipped software

• Mario, the founder of Pi, had built his own ‘Claude for Chrome’ extension and has finally open-sourced it (I’ve been pestering him for a while to do this). Now you can clone it and customise your own Chrome-agent (I will be).

  • There is also an AMA with Mario here (he joins at ~38min mark), and before he shows up, Daniel from Sentry covers his AI coding journey, from copy-paste to personal programming.

• Cal (the better version of Calendly) launched its own agent and skills. The agent is like an EA that can handle scheduling and meetings for you. The skills are to plug into your own agent. I’m not that busy, so don’t use any scheduling agents...but another example of the vertical agent play.

• HubSpot are getting into the vibecoding game? With HubCode you can basically create apps on top of your HubSpot data. And it’s a waitlist, but you can see the demo here. Dharmesh reads this newsletter - my feedback is the demo looks too clunky! Too many things to press/connect/navigate to.

  • SaaS wars are simply every SaaS company trying to cover any surface area their customers may want. Airtable added automations, Zapier added tables, and now HubSpot is adding agents. Lots of these platforms treading on one another.

Dev dishes

• People keep asking ‘where’s cursor for video editing’, and this is one such attempt - loop.

• Andrew Ng’s new project context-hub lets you fetch curated documentation and annotate it for future use. In its latest release, it adds the ability to send feedback to docs authors with up/down ratings with optional labels. So your usage improves them for everyone and creates a new Stack Overflow for AI coding agents.

• A terminal-first twitter CLI where you can get for-you and following feeds, bookmarks, etc.

• If you saw Karpathy’s ‘autoresearch’ last week, it’s essentially a self-improving system designed to help your agents succeed on any task. You give it a task, and it works its way through it, reviews what it did, and tries other iterations to improve. So it was only a matter of time before others made more general-purpose versions.

  • speaking of this loop...a 17yr old, Austin, created a system to teach simulated students Advanced Placement (math?). The first batch scored ~45th percentile, 2 weeks later, ~80th percentile. They were ONLY taught basic knowledge and comprehension.

• Agents need access to up-to-date documentation to be effective. Context7 has been that tool for a lot of folks via their MCP but they just released their CLI.

• Using OpenClaw? There’s now a memory plugin built on top of Shopify CEO’s QMD tool. I’ve tried creating my own, but still haven’t got memory right, so I’m definitely going to test this.

• Mistral launched Mistral Small 4 - they combined their different model lineups in a single general model with this one. It can do decent coding (Devstral), it’s multimodal (Pixtral) and uses reasoning (like Magistral).

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

this is an example of a new format i like the idea of… the original of this email (in our current format) is here. please vote at the bottom of this post which you prefer.

[this top section is where i’ll talk about things on my mind. sometimes it’ll be tool comparisons, sometimes explainers]

A founder I’m working with asked me yesterday if I’d read Sequoia’s post; Services, the new software. I had when it first came out, but I re-read it. TLDR... AI companies that sell outcomes (autopilots) rather than tools (copilots) will capture the far larger services budget.

Example: Harvey started selling copilot as a tool to law firms but is moving towards completed contracts and filings as an outcome.

The best strategy is to start with outsourced, intelligence-heavy tasks as a wedge before expanding into higher-judgement work.

It made me think of the vertical agents that we are seeing today, which effectively start as custom skills for one task. It could be documentation writing, animated videos, explainer visualisations, email triaging, sales reporting, etc.

We are seeing it over and over again now with agents for CFOs, CEOs, CMOs and so on. We will see it for every profession and job level there is.

This was on my mind as I saw the Manicule launch - “an AI native technical documentation studio”. They have packaged up an agent+human services into something that feels like consulting/contract work. I think it’s smart (not just because it links with the Sequoia post I promise).

I think the packaging is what’s smart (tbd on the job-to-be-done they are solving - won’t agents just write great agent docs for agents to read?).

Another founder I spoke to yesterday (I think I’m going to invest) is working on specification and planning for development workflows. We’ve all heard ‘you should use plan mode before coding’ and the more I see people build/fork their own orchestration platforms the more planning feels like the only step to truly matter.

If you could have your team (+ agents) all collaborating on the specification and spending time thinking about tradeoffs, implementation details and the like, the agent should be able to execute that.

All of the grunt work is done in planning. And when code is up for review, it’s much easier to read and make sense of the spec that created the code than trawling through the code itself.

And broadly, code review feels like it could be a thing of the past when models hit a certain threshold. tbd.

Subscribe now

Chewing on

• Codex now has 2M+ weekly active users, and OpenAI API use is 20% up since GPT-5.4 was released. Subagents are also now live in both the Codex app and CLI

• Manus (recently acquired by Meta) launched a desktop app, My Computer, to compete with Codex/Claude Code/OpenClaw, etc.

• Jensen Huang says Nvidia expects to generate $1T+ in sales from its flagship AI chips through the end of 2027. They also released NemoClaw

• Claude’s 1M context window is now generally available.

• What is agentic engineering?

• Travis Kalanick of Uber fame is back with a bang. TBPN video interview

• “The Return of Travis Kalanick: Fact & Fluff!”

• The socio-cultural importance of the Claw movement: Lobster Boil

• Productive individuals don’t make productive firms

• Vibecoding is my passion

• Nebula. The best packaged OpenClaw.

• Customise your own ‘Claude for Chrome’ extension - AMA with Mario here (~38min mark)

• Cal (the better version of Calendly) launched its own agent and skills.

• HubSpot launched HubCode - create apps on top of your HubSpot data.

• Andrew Ng’s new project context-hub lets you fetch curated documentation and annotate it for future use.

• A terminal-first twitter CLI where you can get for-you and following feeds, bookmarks, etc.

• more general-purpose version of Karpathy’s ‘autoresearch’

• Context7 just released their CLI for code docs

• a memory plugin for OpenClaw

Digestif

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ve been thinking about doing this for weeks but I like to really ‘see’ what the end output looks like before I run with it.

But that’s just procrastinating.

So I told myself I can’t open my new MacBook until i’ve sent this 🥹.

I’d appreciate feedback if you like this style of email and what you build with it!

Subscribe now

What did I build this week?

1. Become a builder.

1.3k people signed up for this workshop I hosted last week [i’ll do more]. But Codex crapped out on me during it (hence the new MacBook). I wanted to put together a cookbook to go through everything.

It just ended up as a step-by-step tutorial. It’s boring. Are you going to read one screen then switch to your tool and do it? maybe.

Instead, I’ve been working on an interactive cookbook you give to your agent and it teaches you as you’re building.

At the end, you’ll have built and deployed your own site with all the new concepts you covered whilst building it.

It’s been hard to get this cookbook right, so lets count this as alpha0.1. Please let me know how it went for you, what your site looks like, where it fell short etc and I’ll improve it.

What do to:

• Open Codex/Claude Code desktop app

• Create a new project folder

• Open a chat session in that folder

• Copy this url (the instructions) into your agent, hit enter:

https://gists.sh/bentossell/a4e5e7048e8a355ec56cf3db86169ae2

• You can choose ‘Full Access’ on Codex and ‘Bypass permissions’ on Claude if you feel comfortable (this project just creates a new website for you). Alternatively accept permissions as you go.

I recommend highly reading the agents output, look at what it was thinking in between your prompts.

Fill your site up with any concepts you don’t know and share them, I’d love to see.

Disclaimer: Codex may produce uglier designs than Claude.

2. Visualise skill.

One issue from the above cookbook was visualisations. I think it’s really helpful when learning about code systems.

All my attempts looked like 💩 and then Claude shipped their visualisations yesterday. Good timing.

So I reverse-engineered it and released it as a skill you can add to any agent. Codex still has poor design taste but it’s much better with the skill than without, trust me!

This is my first GitHub project to get over 200 stars!

Just give the link to your agent and say ‘install this skill’.

3. Ben’s Bites Cookbook site

A redesign, again.

The previous cookbook site had lots of dead weight from older versions so I wanted to start fresh.

Code is basically free nowadays after all!

It’s definitely not finished but in a decent place. This is where I want to upload a bunch of helpful docs to help you build stuff and see a breakdown of how I build stuff.

Still a wip! Not live yet. Needs another design pass - contrast is way off for a start.

What’s in my stack - tools, skills, instructions, models

Models. I always mix them.

• GPT 5.4 XHigh for all ‘proper code’ - new features, new ideas etc

• Opus 4.6 - for planning, research, less-technical tasks, design (always)

CLIs (terminal-based tools)

• Droid for when I want to build something properly (their new missions feature is insane, can run for hours by itself and implement stuff end to end) - I’m an investor in the co

• Pi is my new other favourite child. It’s very fast, and lightweight so your own instructions guide it a lot more than others

Both let you switch from GPT ←→ Claude models (or gemini, etc etc) in one conversation.

I use those in the terminal exclusively. I used Ghostty as my terminal app but now I use Cmux which has Ghostty in it, just has a nice sidebar for organising chats, draggable panels and a built-in browser. I do wish it had an easy way to view my files though - until then, I use Zed for that.

Agent Apps or whatever we’re calling these 3 panel agent interfaces;

• Codex app - really nice user experience, super approachable

• Claude Code/Cowork on the desktop app - I very rarely use these but have this week with some testing. I’m not won over by these yet.

• T3 Code - this is nice, snappy and will support multiple agents but for now just Codex. Until it supports other agents I’ve not been reaching for it over Codex for GPT work.

  • I saw Theo’s video ‘leaking’ a command to get an early version. I didn’t know it’d be open source when released so I installed it and asked gpt 5.4 xhigh to reverse engineer it exactly - it did it no problem!

Skills

• frontend-design from Anthropic [link]

  • It works well but I don’t feel like it should when I read the prompt 😅. I’m just waiting for the ui.sh skill to be released so I can use that (from the Tailwind guys).

• json-render from Vercel [link]

  • This is a great ‘generative ui’ skill that can spin up interfaces suuuuuper fast. I use it to make zapier/n8n canvases of automations I’ve got set up on my Mac-Mini. The team are pushing updates almost every day. I need to play around with it more.

• agent-browser from Vercel

  • My go-to for my agents. Spins up a chrome browser, looks at my site, takes screenshots, navigates, clicks, records the screen etc etc - basically use the browser like a human. There’s a ‘dogfood’ tag which grabs all the errors, and writes a report to fix. I am bumping into it not being able to bypass sites with Cloudflare ‘bot detection’ - like OpenAI. Irony isn’t lost on me.

• react-doctor from aiden [link]

  • This has been great making sure I’m using best-practices when my agents use React (quite often). It slots in when things have been built and tests/checks are happening and it nearly always catches something to fix.

What about skill prompt injection?
It can happen. I’ve not experienced it. Use reputable sources like Skills.sh (from Vercel) or just ask your agent to re-create the skill and check for any security issues. Tools like Codex app have a create-skill skill you can use - just ask the agent.

Other tools

• exe lets you spin up virtual servers really easily, has an in-built agent to help if you get stuck. Overall made it super easy for me to feel comfortable with servers - which I wasn’t previously.

  • You’ll want another server if you have an automation or agent you want ‘always on’. If it’s on your computer, it won’t run if your lid is closed!

• here.now - im always spinning up sites for random ideas or even just to present info nicely so i can view it on the go. this is a free tool to give your sites a custom url in no-time at all.

  • I liked this and the founder so much that I invested this week!

• Vercel. Vercel and Cloudflare are mortal enemies on X. I’ve got half of my deployed sites and domain names on both of these. I want to just pick a default one and Vercel’s edging it for me because I’m using a lot of their tools and skills. But honestly this could change by tomorrow.

• gists.sh - I love tiny tools like this. GitHub has ‘gists’ which are quick ways to have a file on a url you can share or keep private - easily readable by agents. But it’s ugly. This tool makes them super nice to share - which is why I put my interactive cookbook in one.

Tools on my list to tinker with:

• Context7 CLI - docs

• Browserbase Fetch API - scraping sites. Need to see browserbase vs agent-browser too

• Ramp agent card - credit cards for agents

• Replit Agent 4 - shall I do a head to head of vibe coding tools?

• Web to Design - Turn any website into an editable UI.

What’s in my AGENTS.md

An AGENTS.md is a markdown file with instructions that the agent loads into its context at the start of any session.

Claude specifically looks for CLAUDE.md - but I just have mine symlinked to one another - ie if you look at claude.md it shows you the agents.md file. Ask your agent to set that up or to use dotagents

You can also paste these in to Codex/Claude desktop apps.

This is the build ‘loop’ that I’ve added.

Any agent I use follows it (italics are there for you - not included in the file):

• create a /spec/ folder.

  • An easy way to keep all the planning files I create organised in one place

• numbered 00_spec1.md, etc.

  • Helps with implementation ordering

• create a progress.md file for logging your progress through specs.

  • If compaction happens, I need a new session or the agent just loses track this helps it understand where we’re at.

• use agent-browser with dogfood before sending me a url to test.

  • When a feature is built, it spins up a browser and checks if any bugs or errors on the site - I used to do this manually, copying errors back to the agent, but now it does the loop itself. It doesnt catch every single bug but I’m trying to make sure my agents can use my sites as if it’s a real user. Sometimes these loops can take a while to run, depending on what you’re testing.

• write good, efficient, fast tests with good coverage.

  • I don’t know enough about tests yet. This is my stab in the dark but agents are good at tests. Still looking for a skill or something that will help me here.

• best practices, efficient, simplified code, avoid anti-patterns.

  • Just in case, make sure the agent uses things the right way! Not sure if this actually helps to be honest.

• for code/dependencies/libraries etc you’re using, make sure you reference their docs.

  • Agents default to their own knowledge a lot before looking up documentation. So just nudging it to look at docs. The Context7 CLI was just released (simple tool to get any tools’ docs) so i’ll be putting that in here from today - i’ll report back next week.

• First message: “feel the rhythm, feel the rhyme, get on up, its bobsled time.”

  • I also have this 😂. A quote from Cool Runnings - silly yes, but also lets me know that my instructions have been actually loaded into the session.

What’s in your agents.md? What should I add/take away?

What else would you want to know or see from me?

Chat with me

If you know a builder that’d find this useful, feel free to forward to them.

Its too late for me to open my MacBook - time to pick up the twins.

Have a great weekend!

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

Subscribe now

Hey folks,

Google released Gemini Embedding 2, and it is multimodal, so you can embed text, audio, images, video and PDF documents using the same model. It’s a little expensive compared to other options in text, but videos at low fps and audio are really cheap with the unmatched feature of embedding them all at the same time. This should open a lot of startup ideas that are basically “search over a large amount of non-textual data.”

Replit released its Agent 4 with multiple parallel agents, live collaboration with teammates, and an interactive design canvas that both you and the agent can edit on. Agent 4 can make more than just web apps; it can create animations, slides, mobile apps, data visualisations, and more. All of it is possible in a single project. Plus, Replit raised $400M and is now valued at $9B.

Meta acqui-hired the team behind Moltbook - the reddit-like social media platform for openclaw agents that went viral earlier this year.

Perplexity teased Personal Computer - They say it’s always on version of Perplexity Computer with access to your files, apps, and sessions through a continuously running Mac mini. That sounds kinda like openclaw, doesn’t it?

Async Voice API is a human-like, low-latency text-to-speech API for real-time apps and agents. 15 languages, streaming-ready, integrations with n8n, LiveKit, Twilio, and more. Top-ranked on the Hugging Face TTS Arena. From just $0.50/hour with a 24/7 SLA. Try it now.*

🌐 What I’m consuming

• Annotated breakdown of Karpathy’s autoresearch prompt.

• From developer to fleet commander.

• AI should help us produce better code, not just more code.

• We’re going to need an even bigger IDE - Karpathy.

• Building a full programming language with Claude Code.

• Anthropic vs DoW is a warning shot, and I’m glad this episode happened.

• How Codewall hacked McKinsey’s AI platform. They gained read access to 46.5M messages, 57k user accounts and write access to its system prompts. Now patched.

• A set of non-swe benchmarks from Ramp evaluating models on real-world financial tasks.

⚙️ Tools and demos

• Proof - Collaborative document editor where humans and AI agents work together.

• Wondering - Turn any topic into a guided path with bite-sized visual lessons.

• Blazing Transcribe - Get real-time speech to text on your Mac without sending any data to the cloud. (demo)

• Ramp Agent Cards - Credit cards for AI agents with spend limits, merchant controls, and full visibility.

• Upstash Box - The best way to give your AI agents a computer.

• Expo Agent - Build truly native iOS and Android apps from a prompt. From React to SwiftUI to Jetpack Compose.

• Gists.sh - Clean typography, syntax highlighting, dark mode for GitHub Gists.

• Comfy UI now has an App mode to hide away the nodes for your users and Comfy Hub to discover & share community workflows.

• Gemini inside docs, sheets and slides can now do more, like formatting the docs, filling in missing data and editing collaboratively.

• ChatGPT now lets students learn maths and science concepts with interactive visualisations. (but they’re all just sliders)

🥣 Dev Dish

• slopmeter - create a nice, shareable graph to show off your Codex, Claude Code, or OpenCode usage. npx slopmeter@latest

• /btw in Claude Code - Have side chain conversations while Claude is working.

• Mastra remote sandboxes - Give your agent a secure, isolated environment to run untrusted user code.

• OpenUI - 3x faster and 67% fewer tokens than json-render to let AI agents stream UI on demand. (repo)

• twitter-cli - terminal-first CLI to read timelines, bookmarks, and user profiles without API keys.

• Firecrawl CLI - Toolkit for agents to scrape, search, and browse the web.

• Parallel CLI - Allow agents to search, access and extract high-quality data from the open web.

• Fetch API by BrowserBase - Simple, cheap and reliable way to get the page content from a URL.

• /crawl endpoint from Cloudflare - one API call and an entire site crawled while following robots.txt. (also see these other endpoints)

• TADA - open source TTS model from Hume. Comes in 1B (English) and 3B (multilingual) parameters, i.e. possible to run on a mobile phone.

• on my watch: agent-browser-protocol — runanywhere cli

🍦 Afters

• NVIDIA plans to spend $26B over the next 5 years to build the world’s best open-source models. They just released Nemotron 3 Super - 120B params (12B active) model with similar performance to GPT-oss 120B and Qwen 3.5 122B

• Two new interesting benchmarks:

  • PostTrainBench - Measuring how well AI agents can post-train language models

  • RuneBench - Long-horizon goal optimisation across 14 AI coding models inside Runescape.

• The Anthropic Institute - New team from Anthropic with a focus on communicating the impact of AI to the world.

• Runway Labs - A generative AI incubator to explore use cases of AI video and general world models.

• Netflix might pay upto $600M for Ben Affleck’s AI moviemaking company.

• Cursor is reportedly raising at a $50-$60B valuation.

Enjoy this newsletter? Forward it to a friend.

Share

That’s it for today. Feel free to comment and share your thoughts. 👋

• Find me on X, Linkedin, or Instagram

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for March? Last few slots available

Subscribe now

Hey folks,

The ‘become a builder’ workshop last week went well-ish 😊 (Codex crapped out on us). The recording is available, but I’m working on a thorough guide to cover everything properly (plus the bits we didn’t get to cover). I’m ~50% through it so hope to have it out this week.

Also, Factory is hosting a hackathon this thursday, everyone gets 200M tokens, and a mac mini is on the line.

OpenAI released GPT 5.4 in “thinking” and “pro” variants. It brings the coding power of GPT-5.3-Codex to the main model series, with better vision, tool use efficiency and a context window of 1M tokens. It’s now much better at computer use (see demo) and financial tasks. It’s also a bit more expensive vs GPT-5.2 ($1.75/$14 → $2.5/$15 per million input/output tokens). OpenAI expects to keep this naming and capacity difference between instant models (GPT-5.3 Instant) and reasoning models moving forward.

More from OpenAI:

• ChatGPT for Excel - An extension to use ChatGPT in a sidebar right from your workbooks.

• Codex Security, an AI app security agent evolved from Project Aardvark—free for a month to Enterprise customers.

• Codex for Open Source - A program for open-source maintainers, giving them 6 months of ChatGPT Pro with Codex, conditional access to Codex Security and API credits.

It’s also acquiring Promptfoo, an open-source AI security testing tool (popular among Fortune 500, stays OSS).

New built-in skill in Claude Code - /loop lets you schedule recurring tasks in a single session, for up to 3 days at a time. Plus, you can now schedule tasks using Claude Code Desktop - these tasks run regularly as long as your computer is awake. They also launched a community ambassadors program for Claude.

For enterprises, Anthropic released Code Review by Claude and Claude Marketplace. The review tool uses a team of agents to review every PR and, on average, costs $15-25 per review. The marketplace lets enterprises consolidate their AI spending by using their Anthropic commitments to pay for other AI apps like GitLab, Harvey, Replit, etc.

Karpathy released autoresearch — agents autonomously iterate on LLM training code. Ran 2 days on 8xH100, found 20 real improvements with an 11% speedup. 630 lines, single-GPU, open source. I assume this approach of agents coming up with ideas and implementing them will see much more activity this year.

Yann LeCun, Meta’s ex-Chief AI Scientist, along with other researchers, has raised over $1B at a $3.5B valuation for their new startup, Advanced Machine Intelligence (AMI Labs). They are already operating from Paris, New York, Montreal and Singapore with a strong focus on world models and research that goes beyond LLMs.

Go stackless and get back to selling. Remember when selling meant talking to people? Before the tab-switching and endless sync errors. Reevo brings it all back to one platform. Prospecting, calls, pipeline, and reporting all in a single tab. From prospect to close. Go Stackless. reevo.ai*

🌐 What I’m consuming

• Cursor’s third era - Cloud agents have overtaken tab autocomplete in the IDE.

• a16z’s sixth edition of Top 100 consumer AI apps.

• Why is everyone in AI talking about filesystems?

• I was a 10x engineer. Now I’m useless.

• Building for trillions of agents - They will need their own infra, access to files, identities, while maintaining security, compliance, and governance.

• How OpenAI uses skills to maintain open-source repos for Agents SDK.

• The next $1T company will be a software company masquerading as a services firm.

• Using claude code as the chief of staff for a boutique consultancy.

⚙️ Tools and demos

• Cursor Automations - Build always-on agents. Run them on a schedule or use events (like Slack messages) as a trigger.

• T3 Code - Desktop app to use Codex CLI (alternative to Codex app). nice and smooth to use, still feels alpha though (because it is).

• Handles by here.now - Personalised sub-domains for everything you publish with your agent.

• Copilot Cowork - Handoff tasks to agents with the ability to work across your Microsoft 365 apps.

• Air by JetBrains - Agentic dev environment built for working with agents from different vendors.

• Clawcard - A real inbox, a phone number, and a credit card your agents can’t abuse.

• 21st Agents - Infra for adding agents to your app—runtime, sandboxing, billing, UI, streaming and more. Also see: Terminal Use (very similar, YC W26).

• Code review tools:

  • Warden by Sentry - Set of skills to review every PR on your codebase.

  • Vet by Imbue - Fast and local code review tool to make sure the agent followed your instructions.

  • OpenReview - Open-source, self-hosted AI code review bot powered by the Vercel AI Cloud.

🥣 Dev Dish

• Notchi - Cute little Tamagotchi that lives in your notch. It cries when you yell at claude and gets happy when you praise it.

• Context Hub - An open tool that gives your coding agent the up-to-date API documentation it needs. (read more)

• Agent Safehouse - macOS-native sandboxing for local agents.

• Flue by Astro - A framework to build sandboxed AI agents and CI workflows.

• slacrawl - Get your Slack data locally with or without API keys.

• claude-replay - Turn claude code session transcripts into self-contained, embeddable HTML replays.

• executor - Local-first execution environment for AI agents. (read more)

• agent-coworker - Agent backend that you can use from a terminal or a desktop app.

• agent-kanban - VS Code extension that provides an integrated kanban board to manage coding agent tasks.

• Fractals - A tool to break down tasks into subtasks on repeat, let agents complete them and manage the entire process.

• Uithub is now open-source. Turn GitHub repos into LLM-ready context.

• shadcn/cli v4 - comes with skills, presets, dry-run, monorepo and more.

• Experimental UI to fork convos and explore side tangents without interrupting the main thread. (read more)

• An agent skill to help you write smarter, simpler, and more modern SwiftUI.

• Making OpenClaw and Codex app talk to each other using ACP.

🍦 Afters

• MultiGen - new research from Google and Stanford to make level design possible for “generated” multiplayer games.

• Opus helped the Mozilla team find 22 vulnerabilities in Firefox in just two weeks.

• PinchBench - ranking the models based on tasks completed successfully on an OpenClaw setup.

• Databricks’s research team trained KARL - Knowledge Agents via Reinforcement Learning to create faster and low-cost alternatives to frontier models for document-centric tasks. (tech report)

• Anthropic is suing the DoD to block its supply chain risk designation, calling it unlawful. Meanwhile, the White House is preparing an executive order to formally ban federal agencies from using Anthropic’s tools.

• OpenAI’s head of robotics, Caitlin Kalinowski, resigned, citing concerns with surveillance/weapons concerns after the DOD contract.

Enjoy this newsletter? Forward it to a friend.

Share

That’s it for today. Feel free to comment and share your thoughts. 👋

• Find me on X, Linkedin, or Instagram

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for March? Last few slots available

Subscribe now

Hey folks,

omg…1000+ of you signed up for the free ‘become a builder’ workshop i’m hosting at 3pm UK (7am PT). So I’m doing it as a YouTube Live stream. Last chance if you want to join - it’ll be recorded if you can’t make it.

Google released Gemini 3.1 Flash Lite - A fast model with better than Haiku 4.5 performance on benchmarks. But this bump in performance comes with a price increase ($0.10/$0.40 to $0.25/$1.50). At this price point, open-source models like Minimax M2.5 make a lot of sense for developers. They would give a much stronger performance (although at the cost of Flash Lite’s insane speed).

OpenAI also updated its default model in ChatGPT to GPT-5.3-Instant. A lot of improvement in the model is around its behaviour. Lesser halunciation, refusals & disclaimers plus better web search usage and writing. Remember, this is the model most people use when they go to ChatGPT. Also, Codex is now on Windows as well.

Google Workspace released a CLI for Drive, Gmail, Calendar, Sheets, Docs, Chat, Admin, and more. It’s really well built with a focus on agents. One of the team members working on it also wrote a blog on rewriting your CLI for agents.

The Information reports that OpenAI is building an internal alternative to GitHub. OpenAI’s browser Atlas came 11 months after The Information reported on it first. So expect this to take some time. It is also talking with The Trade Desk to put ads in ChatGPT and maybe planning an IPO as it hires a law firm, and Jensen Huang kinda leaked it. OpenAI’s ARR is now about $25B, only a little bit ahead of Anthropic’s $19B.

Voice Agents need speed you trust. Two seconds of latency kills conversations. Speechmatics delivers partials <250ms and finals ~300ms, built real-time first, not batch retrofitted. Fewer errors at conversational speed. 55+ languages. LiveKit, Pipecat, Vapi ready. 👉 Start with $200 free credits.*

🌐 What I’m consuming

• Top SaaS vendors on Ramp (March 2026) - Anthropic and Granola are growing the fastest, with Paper (design tool) and many “hosting for agents” products in the trending list.

• 3D as code - half of this post went over my head, but (I think) it makes the claim that 3D models should approach generation as creating structured code for 3D formats vs say generating the next video frame.

• Agent Experience - Practical reference to the patterns, surfaces, and design principles behind AI agents. (open source repo)

• Turning my Twitter archive (36,000+ tweets) into a feature-rich library to write a book.

• Building an AI-native word processor using Cursor’s long-running agents.

• Rebuilding MacOS for the web on Replit.

• Why did someone choose Droid as their daily driver?

• A practical guide to community-led growth for AI startups.

• Wasteland - The sequel to Gas Town.

• A dozen or so principles for the zen of AI coding.

• Software isn’t dying, but it is becoming more honest.

⚙️ Tools and demos

• Glaze by Raycast - Create real desktop apps with features like file access, keyboard shortcuts and more. In private beta now.

• Glide lets you create apps from a spreadsheet. It’s soon moving to AI from its no-code roots.

• Paperclip - Hire AI employees, set goals, and your business runs itself. (github)

• ASC11 - ASCII art editor to animate images, videos and live cams. Preview with HTML and export to JS.

• Modem - Your dev team’s auto-triage PM.

• Nen - All the infra you need to build reliable computer use agents.

• moss - Notes app that organises your work and connects your thoughts.

• Exa Deep - Agents that run in a loop to gather all the information for your search.

• Defuddle - Get the main content of a page in Markdown. Just add deffuddle.md at the beginning of the URL. Use it via curl, Skills or the CLI.

• Tools that let you run multiple parallel agents for coding:

  • Polyscope - Dev environment for running parallel agents with built-in preview browser and more.

  • Moment - A collaborative workspace to run your entire business out of a Markdown file with code at the centre of it. (read more)

  • Superset - Turbocharged terminal that allows you to run many CLI coding agents.

• Design with AI tools:

  • Refero MCP - Connect your agent to a curated library of real product interfaces and user flows.

  • Kodo - Fully editable designs, generated by chatting.

  • /playground - A design canvas for your code.

🥣 Dev Dish

• yt-browse - A YouTube channel search TUI. Filter, sort results by views/duration/date, use regex, order results, and ignore shorts.

• WebPage by Apple - they don’t call it a headless browser, but it’s just that for native apps.

• Hemingway - Turn your dev site into a copy editing surface with changes that sync to your codebase.

• Agents UI by LiveKit - Component library for building polished frontends for voice agents.

• Symphony - an experimental repo from OpenAI that turns project work into isolated, autonomous implementation runs.

• Handoff in the Codex app - a simpler way to move a thread between Local and Worktree.

• eval-skills - A set of skills for AI product evals (read more)

• 30+ skills to make your agent really understand developer marketing.

• Cursor is now available in JetBrains IDEs through the Agent Client Protocol, and it supports MCP apps now.

🍦 Afters

• A new device, Spectre I, claims that it can block unwanted audio recordings of you.

• Mintlify acquired Helicone, an “observability for LLMs” startup.

• Meta is planning a feature called “Name Tag” - Point your glasses, and the AI tells you exactly who they are. It is also building a new applied AI team.

• Qlabs is running an open experiment to train better models with limited data, given a large amount of compute.

• RevenueCat is hiring an agent (at $10k/month) as a developer advocate.

Enjoy this newsletter? Forward it to a friend.

Share

That’s it for today. Feel free to comment and share your thoughts. 👋

• Find me on X, Linkedin, or Instagram

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for March? Last few slots available

Subscribe now

Hey folks,

I’m hosting a free workshop on Thursday (will be recorded) - Become a builder. Aimed at less-technical folks who want to build apps, automations and agents with AI tools. If you’re drowning in content, tool choices and don’t quite know the system to get to the next level, this workshop is made for you. I’ll reveal how I reverse engineer business ideas, build from scratch, set up agent automations and consistently ship new things. Sign up for free here.

Last weekend was not normal. The US government’s Department of War (DoW) labelled Anthropic a supply chain risk (here’s what that means) because they did not agree on Claude’s use for autonomous weapons and mass surveillance. Meanwhile, OpenAI did manage to reach an agreement with the DoW, claiming that it still holds onto the same redlines as Anthropic. It did not, leading to an AMA with the OpenAI team and an update to the said agreement. OpenAI is also publicly asking the DoW to take back the risk label on Anthropic.

This is far from solved, but OpenAI has already taken clear damage to its public image. Anthropic did a cheeky little play of adding an import memory button to Claude for people moving over from ChatGPT. But is that enough? Because if the US administration has its way, Anthropic has much larger concerns. Ben Thompson, as usual, goes into the political details of Anthropic and Alignment.

That doesn’t stop them from shipping, though.

Claude Code has an auto-memory feature now - it learns across sessions and then remembers it when you start a new one. Also, voice mode in Claude Code is rolling out to users. Type /voice and then hold space to use push-to-talk. Your messages get transcribed, and it doesn’t cost you anything, not even your tokens from your subscription.

OpenAI has raised $110B at $730B valuation. This brings Amazon as an investor, and OpenAI will use the Trainium chips to serve some of its products. Some stats from this post: Codex now has 1.6M weekly users, and there are 9M business customers and 50M paying consumers for ChatGPT.

Entire CLI now supports Droid - Share your sessions, chat logs, prompts, and more to keep the context of your work.

Incredible is the fastest way (literally) to work on your computer. Press a key, speak naturally & an always-on AI assistant helps you write + act across all your apps natively. Real voice-controlled 'vibe computing', making you absurdly productive. Invite-only. First 50 readers get access codes.*

🌐 What I’m consuming

• What devtools does Claude Code choose when asked to build something? Stripe, Shadcn, and Vercel have a near monopoly, with newer models choosing newer tools.

• Lessons from building Claude Code - How did the “askuserquestion” tool come to be?

• Background Agents - The current state and the promise of autonomous work across the entire SDLC.

• Hitting $1M ARR in weeks with Polsia.

• How I built a sub-500ms latency voice agent from scratch.

• How I actually practice law with AI in 2026.

• The only way to save SaaS - A case study of Intercom.

• How I run 4–8 parallel coding agents with tmux and Markdown specs.

• A Swedish publication claims that recordings from Meta’s smart glasses end up in front of human annotators with bank details, sexual content and more present in them.

⚙️ Tools and demos

• Have a business idea? With a simple prompt, GoDaddy Airo AI Builder creates launch-ready websites and apps in minutes.*

• Maestro - Teaching partner for educators that adapts to your students and how you teach.

• Readout - Fully native macOS app for a real-time overview of your dev environment and Claude Code configs.

• Instant Apps by Tasklet - Your agent can now build live, interactive apps for you on-demand.

• Mintlify now lets you create docs for any GitHub repo, auto-generated from the source code. (read more)

• iocaihost - no-account website hosting for agents.

🥣 Dev Dish

• pplx-embed - SOTA embedding models from Perplexity trained for large-scale retrieval.

• evlog - One log per request with full context ofwide events & structured errors for TypeScript.

• Evolver - Optimise your code and prompts with a Darwinian approach. (read more)

• recall - A skill to let claude code search over all your conversations from past sessions.

• noodle - Skill scheduler with self-healing and backlog sync. (read more)

• agentcard.sh - Prepaid virtual Visa cards for AI agents.

• actors.dev - Give agents an email address and the ability to make phone calls.

🍦 Afters

• Burger King deployed Patty to employee headsets - a chatbot that helps them with meal prep and friendliness.

• Anthropic is giving 6 months of free Claude Max 20x to open source maintainers and core contributors. Apply here.

Enjoy this newsletter? Forward it to a friend.

Share

That’s it for today. Feel free to comment and share your thoughts. 👋

• Find me on X, Linkedin, or Instagram

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for March? Last few slots available