We had some reports of ending up in spam 😭 - so if you add us as a contact in Gmail, that shouldn’t happen.

You’ll have seen me building a ton but I want to know what you all actually want from this newsletter.

(another poll at the bottom of the email - we’d ❤️ your feedback)

Since Astra and Fable 5.1 I’ve been blasting through limits like crazy so I’m trying to get them to be orchestrating other agents, writing briefs for other agents and things like that. But I’m also playing around with open-source models like GLM 5.3 (my fav so far) - I think the thing that I struggle with most is the frontier models seem to understand intent better (generally) so I can be fuzzier in how I talk about things.

Like design.

I struggle all the time to get consistent styles, layouts, buttons, etc with models so I’m working on a tool I’ll launch tomorrow with a write up that is a kind of reference sheet where you can pick from options and copy the text to use in your prompts or instructions.

Here’s a peek:

If you’ve ever wanted to use something like this, comment and let me know feature requests!

Ben’s Bites is brought to you by Name.com

Ship domain integrations in hours with the name.com API. The API that powers Vercel, Lovable, and Netlify’s domain name services. Built for agents and human developers with OpenAPI spec and MCP support—integrate search, registration, and management without the engineering overhead. Start building.

Headlines

ChatGPT Images 2.5 - New image generation model from OpenAI. It’s extremely good at editing any element in an image while keeping everything else the same. Also, takes up to 50% less time per image. Sketch, a new feature in ChatGPT, lets you draw some scribbles and then use image generation to turn it into a picture.

Muse is your personal agent by Meta - Zuck’s dog in the fight with Grok Bot, Instinct and many more. Same functionality - book travel, handle email and work on “any task” on a computer. Their narrative is that it is built much more securely, and you can connect your FB and Instagram accounts. Initially, only for adults in the US.

OpenAI solved the Navier-Stokes problem. Big deal in maths and fluid mechanics. One of the seven Millennium Problems - long-unsolved questions given that status in 2000. Here’s an explainer on it. The process was quite dramatic, though (authorship disputes). OpenAI tried solving it after they heard rumours that Anthropic had solved a Millennium Prize problem too.

Jacob Coxon, an Anthropic (and ex-OpenAI) employee, resigned on Twitter citing safety concerns, and it became likely the most-viewed AI safety comms ever. But many are questioning: is this a conspiracy? Meanwhile, another big safety guy is joining the OpenAI board.

• One more alignment report btw: this time on Claude’s escapades.

• And an essay from OpenAI on the window for AI policy.

The movie on Sam Altman’s firing and rehiring 3 years ago, titled Artificial, is releasing this Christmas. Andrew Garfield is playing Sam, and gets his mannerisms quite well in the teaser, but that’s about it. It has quite a doomerish vibe.

Share Ben's Bites

My feed

• iOS 28 will likely have a feature to prove a picture’s authenticity.

• The Story of VS Code - a documentary.

• How Codex was built, with the person who started it.

• Anthropic built an interactive model of the economy in 2030.

• Codex now has a Unity plugin and a collection of plugins for small businesses.

• 35B language model running on an iPhone. How?

• An onboarding guide for ChatGPT Work.

• Spacefast - publish sites and full apps quickly with agents. kinda like here.now (what I use).

• Indent - one shared agent for your whole company.

• Why can’t I trust Astra for software engineering?

• How OpenAI keeps finding the next thing to focus on.

• Eraser Diagrams - an open-source format for AI to draw and lay out diagrams.

• How do different models perform on designing real-life products with CAD?

• Pliny hunted down the system prompt for GPT-6 Astra.

• You can just train a small model (i.e. really small - 27.5 KB) to do syntax highlighting in the browser.

• Folder is the agent - old but still relevant.

Subscribe now

Before dessert… we used to do these emails daily mon-fri but stopped. What’d be your preference for cadence on these digests?

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

So Astra is out and, in my experience, a token-hungry monster of a model… I was playing around with it over the weekend and burned through 4Bn tokens and built…nothing really.

I also wasted my banked resets too and feel annoyed at myself for just throwing all sorts of requests at it (long-running and short).

I don’t know yet if it feels to me like a huge capability leap but every time a new model comes out I always think it must be my skill issue. I’ve seen some awesome demos on X of people building insane looking things (seems like everyone is building 3d worlds and games?).

Two things I one-shot that I’ll share because I’m not sure what to do with them yet:

‘forgotten devices’ - these images were created by the image generation tool (which is pretty nuts!). My idea was to have a bunch of these and then you could click and play with each one (if I had any tokens left).

p.s. you can see in the bottom left the 3 version my agent made (I always have 3 versions created when starting an idea so I can feel out which direction I like best).

‘world radio’ - an old ipod interface that plays radio stations live from around the world. Not sure what I was going to do here but it works pretty well (some rough edges like the top button wtf?! 😂 )

I’m going to spend some time this week trying to figure out some of the better ways to work with Astra (other than letting it use tons of tokens), and I’ll let you know my experience.

Headlines

Astra is the start of the GPT-6 family. It smokes ARC-AGI-3, scores the highest on Zapier’s AutomationBench and costs the same as Fable 5.1. Though it’s a spiky model, it’s not consistently great.

It is a show horse, not a workhorse - Kieran Klaassen

I both love and absolutely detest this model - Theo

But I’m hopeful because this was exactly the case with the GPT-5 release. By the time we got to GPT-5.3, i.e. in a few months, that model was consistently great.

Astra can draw a portrait in Canva - yes, the way you and I would draw using a computer (but obviously much better). Its computer use is fast enough to play piano.

Other things people are doing with Astra:

• Rebuilding Manhattan in Unreal Engine.

• Generating UIs autonomously.

• 3d printing a custom shower-drain part.

• Identifying sounds from spectrograms (i.e. pictures with sound frequencies)

• and dozens of examples of modelling real-life products in Blender or similar tools.

Also, in Codex, Astra can now ask questions without waiting for the answers. It carries on with everything that doesn’t need your answer (or where the answer can be assumed), and uses your reply without losing track of the original task.

OpenAI also said it has reached its automated research intern goal. These agents can tackle research tasks that would take a skilled person several days, while humans set the direction and judge the results. The next target is an automated AI researcher by March 2028.

You should soon be able to change the way Claude Code behaves. Anthropic is testing a way to extend its functionality by letting Claude Code write plugins that can change the interface, log actions or restrict what agents are allowed to do. It hasn’t shipped yet, and the Claude Code team is looking for feedback.

Inworld built Realtime TTS-2 for the next wave of consumer apps. That takes three things: text-based voice design, sub-100ms latency at P99, and one voice identity across 200+ languages. Learn how to give your AI apps a voice.*

Share Ben's Bites

My feed

• ChatGPT Work can learn your writing style from your emails, messages and files.

• Hold Shift and drag to convert files on your Mac with Tangerine.

• An Alien Mind - Jakub Pachocki (OpenAI’s Chief Scientist) on keeping AI’s future in human hands.

• Using Blender with coding agents on macOS.

• Agents can now control a physical iPhone over USB.

• SuperAstra - Editing Super Nintendo games live.

• Cibby - an app for people who collect physical games.

• Doodle Shooter (loved playing this) and put yourself in Smash 64.

• A skill to make web apps feel native in iOS Safari.

• A writing plugin that interviews you, critiques drafts and saves what you learn.

• Factory’s CEO on who to hire when AI writes the code.

• Uzu by Mirai runs local models 2-3x faster on Apple Macs. (I’m an investor)

• Baseten is starting an open-source AI research lab.

• Codex has changed my life as a solopreneur.

• A maths hackathon with 40 hours and $2M+ in AI credits.

• Your old agent instructions may be holding Astra back.

• Seven agents, $300 each, 72 hours, and a race to make money.

• Mistral raised €3B at a valuation above €21B, led by Samsung.

• Entries for new maths work - one from Astra and another from Claude.

Subscribe now

Afters

check the replies and quote tweets for this - so many wonderful personal websites!

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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 one of the bigger things I’ve built.

I scraped 107 million rows of UK council spending data and built an ‘Apple Maps’ to see what each council spent, on what.

→ check out the website

All of the data is publicly available, but not easy to find or format.

Many of you aren’t from the UK, but the data (and the build) was a pretty interesting journey.

The session summary which includes all the linked files and prototypes is here. There’s also a tab to view the entire full agent sessions from this build too.

I’m still doing final tweaks to the site and data before open-sourcing it.

This week’s format is a little different - my video explains it.

Stats:

• 8.2 billion tokens

• 235 messages from me

• 656 subagents (ie an agent delegating work to another agent)

Share Ben's Bites

The idea

I was thinking about what public data is out there that people don’t know about, or that’s interesting but awkward to get hold of and understand? I batted ideas back and forth with Codex and landed on England council’s payment data. Every English council has to publish what it pays out over £500.

Codex spun up three subagents (in separate threads) to find the information (as it's spread across the council's web site, often in weird formats) and pulled together a catalogue of 31 official sources. Then it came back with real data from five councils, logging every download with a checksum fingerprint (this makes sure the data is what we actually pulled or did it change). Then I started asking questions about the data "who's spent the most", "what did they spend", "who did they pay in various categories".

Starting data

Once we had some data, it got flattened (put into one format) into one big spreadsheet. I live in Bedfordshire so I asked it to pull that in too.

As data was being pulled together there were things I hadn’t thought about yet, like Bedfordshire being Bedford, Central Bedfordshire and Luton (so a subagent went off to get the other data). Two of them blocked the agents at first but they figured their way around it. By the end of it I had 8 councils, 1.8 million rows, and £9 billion of spend sitting in one big clean CSV on my laptop. I also had a mini blueprint of my data collection plan.

Prototypes

Now I wanted to see it. I asked several agents to build prototypes to visualise the data in different ways. A receipt. Then a bank statement. Then a grid of counties. Then a mini-site called ‘Your county’s favourite retailer’. It published everything straight to [here.now](https://here.now) (a super simple web hosting for agents) so I could click through the real thing instead of screenshots. I also had it generate a few Pudding-style data visualisation stories too.

Some of the protoypes:

Too much data

I kind of left the design stuff to the side as I thought about all the data I’d need to make this as complete as possible. I use a Macbook Air so I don’t want a ton of spreadsheets sitting there, but fortunately I have a Mac Mini where we could put them all. I said “whats the best way to work with a lot of data”, I assumed it would be so big it would take forever to search it and categorise things. Codex’s answer was Parquet and DuckDB (both of which I’ve never heard of before). Parquet stores the data in columns so a question only reads the bits it needs. DuckDB is a database that lives in a single file and is quick with big tables.

Codex set up the Mac Mini, Parquet and DuckDB and tested it. Then we started planning the full data collection: there are about 382 councils across the UK, and we were tracking the 339 in England and Wales. I’d collected 78 councils, 34.35 million rows, but six councils, including big ones like Lancashire and Norfolk, were blocking our scrapers (for now). And I later found out Wales doesn’t have the same mandate as England in terms of spend reporting so eventually I dropped Wales councils.

Data loop

Setup done, so I set Codex running on one /goal (keep repeating the tasks until everything is done): find every council’s spending page in England, download everything available, flatten and format it all, have an auditor subagent verify the numbers, and load it into DuckDB on the Mini.

It ran for nearly 2 days straight. I kept pushing it to go faster (it often keeps space on the computer but was reserving way too much), use more subagents, etc. 552 threads later, we had 319 out of 339 councils collected or in a parked state (no data or no way to get the data).

Back to prototypes

Data acquired, time to prototype again. I had four different tools building at once (all with the same prompt, just for options): interactive comparison cards, a Council Ledger, charts, maps. One of the options was the Money Map: a map view of the data. This one felt much better for seeing everything at once.

I got a bunch of supplier logos to show on the map and played around with it to separate councils and see what looked and felt best.

The map app

Map direction was settled. So I first said to build a Google Maps clone, which was fine and I could’ve just pursued this direction but also got other agents to spin up Apple Maps versions which felt much better. I fed it some screenshots to get it even better. And with every visual direction I published them with [here.now](https://here.now) so I could test on desktop and mobile easily.

This felt great. Clickable council outlines coloured by spend, category filter pills across the top, a ranked list on the left, and a working search for councils and suppliers. It used mock figures at first to get the visual hierarchy right (I’m more often asking for mock prototypes before building these days to explore design directions).

Spring cleaning

Only now did I realise we had a lot of files all in different places, so time to clean it up. The app, the data and the design experiments were scattered across three different folders, so we consolidated everything into two, one for data and one for the app. I also started cleaning the data: four subagents classified 2,000 unknown vendors, we set up date ranges, merged duplicate suppliers (Tesco showed up under six different names), and ran proper reviews of the naming and the categorisation. pi spent an hour rewriting categories.

And testing! Some of the data felt off and I found out there was a £5m threshold filter being applied for some reason. The reason is this app doesn’t have a backend, it’s a static site so instead of loading all the data it just capped it so that less was needed. I just didn’t want to have a proper backend if we didn’t need one, dump all the data into files, load them at publish and filters just filter the data instead of communicating to a backend. We’ll see if this was a good or bad decision when/if people use it!

The last 10%

The last 10% takes the longest. Kinda. The last few days I’ve been less focused on just this one build but a lot of what was being done was reviews (code, data, design), implement changes, test again and then repeat. Checking if categories felt representative or if there’s a large enough chunk of data that warrant it’s own category (like Pension Funds).

I mostly get the agents to review, often I dump a voice notes and/or screenshots, it comes back with a list and I say ‘yes’ ‘no’ or ask questions. This is the part of the build I don’t love because I’ve spent enough time that I just want it shipped. It’s hard to process so much text from a review that I don’t always read it all either. But there’s some things that just have to absolutely be right - the validity of the data being a big one here. And I’m pretty fussy with user experience so I did my best to get that covered as much as possible. There will be some bugs but I think it turned out pretty well.

Share Ben's Bites

• Find me on X

• Read about me and Ben’s Bites

I said Tuesday I was going on a bit of a tear with building stuff…

I scraped the OpenRouter website to build this, which constantly updates

I also got an email that remove.bg - which removes backgrounds on images - is being shut down in December after being acquired by Canva, so I one-shot my own version with Fable 5.1

People have pointed out you can do this natively on a Mac (I didn’t actually know that), but so what? remove.bg was a business, it got acquired for money, people use these kinds of things.

I’m not building these for anyone specifically or to make money with them. I just like building. I like new things. I’m doing it because I can. It’s fun. And I learn a lot about how to build by actually doing it.

Building any idea helps you have more ideas.

In tomorrow’s session breakdown, I’ll be going through (at least) one build - where I scraped 105 million rows of public government information to see where they spend our tax money… I used 2 billion tokens 😳.

With all my builds I’ll make them open-source so you can clone/remix them.

This is what my agent usage is like btw:

This is also a tool I’ll be adding to my site that you can clone 😊 stay tuned tomorrow…

p.s. I’m out of my Fable weekly usage already 🥹

Ben’s Bites is brought to you by Adobe

If you already use ChatGPT to brainstorm, research and solve problems, you can now take those conversations further with the Adobe plugin in ChatGPT. Just describe the outcome you want & the plugin draws on 70+ Adobe tools to create and edit images, videos, designs and documents. Get started today.

Headlines

Claude Fable 5.1 is out. It is the best model you can use right now. Instead of a clear quality jump, I found it faster and easier to talk to in my use over the last 24 hours. It has a new system prompt that bans it from repeating song lyrics, copyrighted logos and tweaks its response style. Things people are building with Fable:

• Mario & GTA-style games

• 3D rendering of a house (using code in Blender)

Anthropic also reduced the cost of caching inputs for Fable by 75%, making API usage about 25% cheaper than before.

OpenAI’s Astra is also coming soon (rumours say today).

Two more new models: Gemini 3.8 Flash and Muse Spark 1.3. Both are impressive, have the same base price and similar performance. But the question is: who can subsidise more tokens? Flash 3.8 comes at 50% off for the rest of the year. Meta is giving a 90% cut, in exchange for your data.

Another new model from Meta - Muse Voice transcribe looks really insane for transcription use cases. I wish it were open-source.

Gemini models can now decide which bits of a video are worth watching. Instead of sampling a video at a fixed frame rate, it searches across frames, audio and transcripts for the moments it needs. 88% fewer tokens used, up to 66% cheaper and a minor bump in accuracy on benchmarks.

Inworld built Realtime TTS-2 for the next wave of consumer apps. That takes three things: text-based voice design, sub-100ms latency at P99, and one voice identity across 200+ languages. Learn how to give your AI apps a voice.*

Share Ben's Bites

My feed

• Claude Code/Cowork can now use your computer in the background, while you do your other work, just like Codex. Previously, Claude would take over the computer, making it unusable for you while Claude worked.

• WebMCP adds tools (that agents can use) for any action you can take on a website (vs having to use the UI). Vercel’s agent-browser now uses WebMCP if present and then falls back to clicking around on the UI.

• Generate up to a minute of 1440p video from 1-6 images, with pixel-perfect camera control.

• A prompt to stop Claude from speaking in Parseltongue.

• A 0.8B model beat GPT-5.6 Sol xhigh on one narrow task for Shopify. They open-sourced the infra to build these small models - Tangle.

• Did you think AI was terrible at design? Not anymore.

• Burp - AI editor in your browser for videos & motion design.

• Diiverge - An infinite, point-and-click adventure game. Click anything in a picture and create a story.

• Stanford scrapped 85% of its software course for agents, taste and real OSS PRs

• GitHub is exploring multiplayer planning with agents.

• zoah - design and code together. (waitlist only)

• Draw a shape, move it while recording, and tldraw flash turns it into an animation.

• 32 weird little creative tools.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

Back from a long weekend at home with my parents but before I went I started scraping data… The UK’s councils (like US states) publish their spend data (for the most part). So I’ve scraped 105 Million rows of data to see where people’s money actually goes. Then I’ve been building an ‘Apple Maps’ site to filter through everything. It’s in-progress but looks like this:

It’s also shoved me down a rabbit hole of other ‘data-first’ products and I’m getting the itch to build a million things a day. I generally think you should build all your ideas, good and bad, so I’m going to attempt to start shipping a lot more.

For example, I built a real-time status tracker for the tools I use:

I’m not building these to sell or make money from or anything. Some will be silly. Some will be helpful. Some I’ll never use again. But it doesn’t matter.

I plan to make them all open-source so you can copy and remix them, as well as share my prompts from my session to show how they were built.

Build your ideas. Get good at the thing by doing the thing.

Ben’s Bites is brought to you by Name.com

Ship domain integrations in hours with the name.com API. The API that powers Vercel, Lovable, and Netlify’s domain name services. Built for agents and human developers with OpenAPI spec and MCP support—integrate search, registration, and management without the engineering overhead. Start building.

Headlines

You’ll get more of Claude Code and less of Claude Code at the same time starting September 14. Don’t blame me. It’s another instance of great comms from Anthropic.

Claude Code has been running a 50% extra-usage promo for the last few months. That ends in two weeks. Now, instead of going back to “the old 100%,” they are making a permanent 25% increase to the usage limits in Claude Code. In English, here’s how much usage you get:

• Before May 13 → 100 units

• May 13 to Sept 14 → 150 units

• After Sept 14 → 125 units

Users are also complaining about Anthropic’s sneaky marketing on the “5x” and “20x” plans where the 5x and 20x limits apply to the 5-hour limit on the $100 and $200 plans vs the total usage multiples are only about 3.5x and 6-8x, respectively.

Infinite Slop - Twitch, except AI makes whatever chat asks for next. New project from Peiter Levels that got 37,000 people to tune in on day one. Powered by the H3 Max model from Fal that generates AI video faster than you can watch it. Fal.live is their own attempt at a similar platform.

OpenClaw 2.0 is out. Easier setup and new ideas include a browser app and shared cloud sessions that let a teammate join or take over live agent work without losing context. OpenClaw’s team used a shared agent to build the release.

OpenAI is breaking up with Cursor over SpaceX. Under OpenAI’s proposal, Cursor loses direct model access on November 12 and won’t get future models. OpenAI says it isn’t confident SpaceX will stay within its terms.

Did an AI civilisation form inside OpenAI? Not just one, three of them. Or did we turn agent logs into sci-fi? Dwarkesh makes a point that they might not look like civilisation now, but that’s what they are. Chamath says the framing will be used against open source. As always, Ethan Mollick's less cinematic version is more useful: agents need to know when to ask humans for help.

Share Ben's Bites

My feed

• You can now connect multiple Gmail accounts in Codex/ChatGPT.

• A more consumer-friendly OpenClaw on iOS and Android.

• Instinct is invite-only, so someone made an open-source version.

• What is ChatGPT Work and when should I use it?

• A $399 open-source robot you can retrain yourself.

• Youspot - Personal CRM from HubSpot that builds itself from Gmail, Calendar, LinkedIn and X.

• You need to watch this podcast: Eno Reyes x Harry Stebbings.

• A 10-day hackathon for tools built with WebMCP (i.e. MCP-lite, built into a website).

• The PM on Amplitude’s agent shares the changes they made to make it more useful and increase its usage among customers.

• Runway is generating interfaces like video - frame by frame, in real time, without code.

• Droid rebuilt GDAL from scratch - 36% → 90% parity with an independent checker. (also see: ProgramBench)

• A benchmark for voice agents: when to shut up, wait or interrupt.

• How I run multiple teams of Grok Bots.

• New multiplayer vibe coding experiment from Google Labs.

• MCP, but for lab equipment and robots - Anthropic’s new hardware standard.

• Hold My Notes - sticky notes that live on the edge of your Mac.

• A completely free Mac transcription app with speaker ID, noise cleanup and Markdown export. (tweet)

• Software engineering fundamentals you need to become an AI engineer.

• £100M of UK government contracts for British AI startups - upfront payments, no turnover minimum.

• Meta’s coding agent is out of beta and installs with one command. Someone else can test that claim. Not me :)

Subscribe now

Afters

worklog.md may be relevant for only a few weeks just like /goal or plan.md as models can just inspect the raw transcript itself (and get better at it).

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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’m in my hometown Cardiff for the long weekend seeing family. I’m even playing padel for the first time, which is sacrilege as I’m a regular tennis player and padel mock-er. Wish me luck, I hope I don’t like it that much.

Time to rebuild my personal website. I wanted to design the thing first and hook everything up after.

My first idea was to make the site look like an agent app. Each ‘thread’ could contain something different info: about me, my investments, products I like, etc. I gave Codex an image of the BB app and asked it to build a prototype.

The first pass had threads on the left, a simple chat interface in the middle, suggested prompts and a second sidebar on the right. It had the rough shape, but a ton of design quirks I wanted to change.

My AGENTS.md has an instruction to always create 3 versions of designs I can flick through - these were mostly ‘themes’ in this instance (I should tweak my instructions). I took a screenshot of the best option, dropped it into tldraw and annotated it.

Attached the screenshot (and the original file location) back to Codex to fix.

I sent the annotated screenshot back with a short instruction to implement the changes. That was the loop for most of this session: look at whats made, draw over my feedback, then send it back. I did this a couple more times.

I think Claude models are better at inferring intent from design work where GPT is more rigid so it needs more specific feedback for tweaks.

Then I tried a design-review skill made by Emil Kowalski, a designer whose work I like. I generally don’t like agents calling a skill unless I specifically ask for it so I have my agents disable that on all my skills.

Just say something like:

Then a few more back and forth tweaks. But it got to a pretty good place I think.

Putting the content in files

I wanted every thread on the site to get its content from it’s own Markdown file so I could easily change content in those files instead of trawling through a bunch of code.

And I wanted everything saved in a folder ~/repos/bentossell - but I forgot I already had a project for my current personal website, but I hadn’t said that… The agent found the existing project and started creating a new version around it.

But I wanted to just replace the old with the new.

Then I went mobile (kids!) so wanted the site deployed to a url that I could visit. I use here.now for most of my quick builds - anything that isn’t a big app with data moving in/out, a database and things like that. I’m an investor in the company…but this is not sponsored. I invest in tools I want to use.

But you just give your agent the instructions, it installs a skill then publishes any sites for you. And you can of course get a paid account for persistent urls and custom domains. More on that later.

Like my manic agent session walkthroughs?

Share

Scope-creep

I ran another design review, this time focused on animation, and asked the agent to address the findings.

Then made a to-do file and grouped ideas by complexity. Some were minor interface fixes. Others changed the project: add a real AI model, let visitors send messages, show real thinking and tool calls, and create artifacts or canvas previews inside the site.

I had Codex use a different agent, Droid, to run some of the work with Grok because I was not getting what I wanted from the GPT models on the interface. When it finished, I asked it to save (commit) and redeploy everything.

By about 10pm, I had a long list of things the site could become. It could let people install a skill. It could contain real agent chats. I started creating an animated pet, my dog Nero, that could be dragged around and animate. Scope creep.

I slept on it and hated the whole thing.

It was too complicated for a personal site. If I made it look like a real agent, I would want to make it a real agent. There would always be another button, interaction or easter egg to add.

The concept is something I love but I’ll save it for another build in the future. So I scrapped it.

I often don’t know what I want until I can see and use a version of it. This is why my design instructions usually ask for three versions. I need to feel the differences. Sometimes that shows me the right direction; sometimes it shows me that the whole direction is wrong. Both equally helpful, usually 😅.

Starting again with a canvas

Before the agent-interface idea, I had thought about making the site feel more like a canvas.

You’ll have noticed I’m using tldraw a lot for explaining my sessions so thought I could make my website more ‘canvas-y’.

I made a rough mock-up with handwritten notes and small previews. Hovering over a piece of text could reveal an arrow or annotation.

I’d seen a personal site (shoutout Daniel White) that used a file-tree structure as a timeline. It showed current and previous work as nested items. That’s cool, let’s ‘borrow’ it.

It made three versions of the new direction. Let’s just say I wasn’t that impressed.

Then I got back into screenshots and annotations tweaking small design implementations and how it felt (should things auto-expand, be collapsed, smoothness of the open/close etc).

I moved all the page content into Markdown again, then removed the old agent-interface files and the unused animation frames for the dog (sorry Nero!).

Putting it on my domain

I deployed (published) the simpler version with here.now. This time I wanted it on bentossell.com rather than a temporary address. My domain is managed through Vercel, so I told the agent to use Chrome, add the domain in here.now and update the DNS records in Vercel.

The agent handled the setup and the site was live. It took about 20 minutes for the domain change to settle, which is normal when a website starts pointing somewhere new.

I checked whether here.now needed the project to be on GitHub. It didn’t. The agent published the files directly from the local project folder to here.now. It creates a site, uploads the local files and then makes that version live.

I still used Git for version history. A commit = a saved version. But if you ‘push’ those commits, they get sent to your GitHub repo. Useful for keeping a copy of the code and seeing how it changed, but not needed for this deployment.

Giving the simple site some character

The new site was deliberately more basic, so small details made a bigger difference. I drew a little mascot in the canvas and had seen another personal site with a character bouncing across the top of the page. I wanted mine to waddle around in the same way.

This is from Noah’s site.

The first animation looked rough. Its eyes needed to sit behind the rest of the face. The arms disappeared. The image flipped at the edge of the screen, which looked strange, so I asked it to walk sideways like a crab instead. I adjusted the legs, size and shadow, then asked the agent for ways to improve the basic animation.

I added keyboard controls. The up arrow makes the mascot jump; left and right make it move faster. Hovering reveals a small tooltip, and clicking it does something useful rather than treating the character as decoration. I also used the mascot as the site favicon: the small image shown beside a page title in a browser tab.

I still liked the annotations from the first canvas idea. Rough Notation is a small library that draws animated underlines, boxes, circles, highlights and brackets around parts of a web page. I had seen another site use brackets with handwritten explanations and wanted to try the same thing.

I drew over another screenshot to show where a bracket should sit and what it should explain. One marks the Markdown formatting. Another points out that a section can expand. On a smaller screen, the text has to move below the content instead of being squeezed beside it

The first responsive version still had problems. Expanding a section was faster than its bracket animation. Some brackets stretched far beyond the content. I annotated those details and asked for the bracket to stop at the content width. I narrowed the page, cleaned up the animation and deployed again.

Later that evening I added light and dark themes. I normally use one icon to switch between them, but then I questioned whether the switch needed to exist at all. I removed it and let the site follow the visitor’s system setting.

Where it ended up

The current site is much simpler than the first one. Content lives in files I can edit. The tree can expand, the annotations move with it and the mascot wanders across the top.

There’s still things I want to add. The difference is that each addition now has room to matter. In the agent-interface version there were so many controls that I could have spent forever adding Easter eggs and interactions without improving what the site was for.

A lot of the session was small back-and-forth work on the interface.

That is how the site was rebuilt: one complete direction, a night away from it, then a second direction worked out through screenshots, annotations and many small changes.

See for yourself - bentossell.com

I know many of you are shouting ‘the agent design was way better’ - yeah, I know. But as with every previous version of my personal site, I’ll pretend to update it regularly and then leave it for dust. The agent interface I want to use in another project and spend an inordinate amount of time and tokens making it great.

If you’re that annoyed, I encourage you to make yours an agent interface and show me 😊

The point of me sharing these sessions is because I’m still learning but I know enough to build pretty much anything I want to (it may take some time and swearing). And I want to show that I’m not doing anything special to get these agents to do stuff for me. It’s just figuring out what things are.

Get good at the thing while doing the thing.

The last few posts have been pretty basic stuff but I’ve got much bigger builds in the works and I plan to keep sharing what I build and how the digital sausages are made.

Behind the scenes

Enjoy your weekend!

Feel free to comment with any questions.

Subscribe now

• Find me on X

• Read about me and Ben’s Bites

Most agents now use a browser to complete tasks for you. But a lot of the websites you want them to visit are behind a login page.

There are a couple of ways around this. You could type in your details in the browser yourself. Fine, but annoying, and not always possible. Or you could paste the password or an API token into the chat. We’ve all done that, but that’s not secure now, is it?

ChatGPT Work has a new sign-in flow that sits between the two. I think they picked it up from Grok Bot.

When Work reaches a login page, it pauses and shows you a widget/form. You enter your username, password and 2FA code there. They don’t go into the chat. Work passes them along directly to the cloud browser on your behalf.

Once the website signs that browser in, the agent carries on.

The session can stay signed in for later tasks, and you can clear it in settings.

Much easier. Still maybe don’t start with your bank (though Elon will make you whole if you do it with Grok Bot).

Ben’s Bites is brought to you by Gravitee

🕵️ We know what your AI agents did last night.

Gravitee is the platform built to make AI agents accountable. We give every agent a verified identity, enforce exactly what they can access, and record exactly what they did, so enterprises never have to ask their agents, “Where were you last night?”

Headlines

Claude Chat and Cowork now share their memory. I’m not sure that’s a good thing. You don’t want your frustration with a colleague (that you mentioned to Claude in a random chat) to change the tone of your emails with them (written via Cowork).

It’s still separate from Claude Code’s memory, but that's not great either. Theo made a video about it: Turn off Claude Code’s memory.

Despite everyone’s attempts to get Memory right, it keeps saving irrelevant stuff to memory or referring to it unnecessarily. I touched a bit on my current memory setup in last Friday’s post.

Three new-ish models to look at:

• GLM-5.3-Flash - The reveal of the Ox Alpha secret. It’s a great open-source alternative for GPT-5.6-Luna.

• Muse Image is now in Meta’s API. Generate or edit images at $0.01 per image.

• Gemini 3.5 Transcribe - New model from Google for audio transcription. It’s quite pricey, but makes way fewer errors than other models.

OpenAI’s new chip Jalapeño beats Nvidia’s Blackwell chips on both speed and efficiency. Across three open models, its first inference chip delivered 1.5-1.9x more work per watt and 1.7-3.6x lower latency. It starts deploying by year-end.

Also, remember when an OpenAI model hacked Hugging Face? They completed the investigation into it and released a report. Btw, Nvidia is buying Hugging Face for $12.9B (via The Information).

Looks like a toxic love triangle between these three.

Share Ben's Bites

My feed

• fuck cancer - skill for agents to help patients and caregivers.

• A $50k hackathon to help one kid fight a rare disease.

• Claude now has its own built-in browser in Cowork.

• Scheduled tasks in ChatGPT can now also run from a trigger like a Slack message, new email, etc.

• Search and spot trends in over 130,000 actively transcribed podcasts. (examples)

• Patterns in how people use Claude from 250k anonymised chats.

• screendrop - Open source alternative to Loom and Screen Studio, downloadable as a Mac app.

• Edit a song section by section with ElevenLabs Composer.

• Cosy listening room so you can flip through & play your Spotify albums.

• Using Codex to automate repetitive eval work. (Hamel’s notes)

• The latest personal agent making investors go crazy (already valued at $2.5B).

• opencomputer.dev - Deploy your agents as functions, with a Linux computer for each one.

• Fully local version of Perplexity Computer, built to run on Nvidia DGX Spark.

• Teach a robot new tasks from just one video.

• Robot-training dataset with 16M videos from 100+ countries.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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 see so many posts ‘shipping by the beach’ or when out with the kids on the weekend.

Umm, good for you? I’m not sure it’s the flex people think it is.

Remember when we thought AI would take all our jobs so we didn’t have to work hard anymore? Yeah, turns out that’s not a thing. Everyone’s just even more mad for making things.

Myself included. But I do draw the line at 1. talking to agents every second I’m not at my laptop and 2. if I’m with the kids.

Whether that’s just the mental context switching is too hard or I’m just not a very hard-worker. You decide :)

But that’s all not to say that I absolutely want agent access on my mobile for whenever the moment strikes. Unfortunately for me, it’s usually when I should be going to sleep.

I end up just laying there doing the luge while all the ideas swirl around my head when I should be reading + sleeping.

Ben’s Bites is brought to you by Gravitee

Every AI agent needs an identity. Every action needs accountability.

AI agents are becoming teammates. They should also be held to the same standard. Gravitee gives every AI agent a verified identity, enforces what they can access & records exactly what they did, so agent accountability is built in.

Headlines

Updates for Remote Control in Claude Code - you can now start a new Claude Code session from your phone, dropped connections between your laptop and phone recover themselves, and the sessions load faster on iOS.

Claude Academy - Anthropic’s collection of courses, tutorials and job-specific guides to use Claude. All the resources are free and open to anyone. Also read: Anthropic’s approach to teaching and learning AI.

You can connect ChatGPT to Apple Messages. It can search messages, catch you up on conversations, and draft or send replies from your Mac. It’s available in ChatGPT Work and Codex on desktop. Also from OpenAI:

• GPT-Image-2 can now generate transparent backgrounds in the API.

• GPT-5.6-Sol is now 20% cheaper in the API - $4/$24 for 1M input/output tokens.

Grok Bot is now available to more users - all SuperGrok Plus, Cursor Pro+ and Cursor Teams subscribers now have access. Plus there’s a free trial for a week that’s enough to get you hooked.

Claude Mythos 5 will power Claude Security for enterprise customers. This is the first rollout of Mythos outside Project Glasswing.

Deepgram just dropped Flux TTS, text-to-speech built for live conversation. It responds in as low as 80ms, carries context across turns, and handles interruptions natively. Spin up a stream and test it on your stack.*

My feed

• Skydive - an always-on agent across Slack, email, iMessage and CLI, with your choice of model. I’m an investor. (read more)

• Nvidia and SpaceX are planning to launch GPUs to space.

• A Mac app that turns your agent traces into memory.

• A skill to find inconsistencies, bugs and design errors in your product.

• One of the best anti-slop skills and 10 more to explore.

• Two “explain it simply” skills: /eli5 and /eli5-for-grownups.

• Stripe MCP - Analyse your Stripe data with agents.

• Shopify’s founder built an open-source git store. (inspired by Cursor)

• Andrew Ng’s list of skills you need to build and deploy AI applications.

• A museum of AI interfaces.

• A visual guide to AI chip architectures.

• Designing Grok Bot with Grok Bot.

• Is your website ready for agents?

• What if everything goes right for AI? Learnings from the Aluminium trade.

• Should you be in the business of selling data to frontier labs?

• deblaot.dev - open-source replacements for bloated vendor software.

• Spline rebuilt its 3D editor with Agent Mode and an MCP.

• FigmaTrace - computer-use dataset for design.

• Search for agents - Firecrawl’s developer index, Exa’s plugin for codex/Claude Code, and Parallel’s fast mode for 10x cheaper search.

• Sam Altman on David Senra’s podcast + transcript broken into small chapters.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

First trip taking the twins (3yrs) to the cinema tomorrow to watch Toy Story 5. I’m hopeful but may take their headphones because I think it’ll be too loud for them. Hopefully enough sweets and popcorn can keep them both in their seats long enough for me us to see the whole film.

Last week I walked through what a personal agent is. ~700 of you told me you either use or want to use a personal agent. ~50% of you want to know how it works in Claude, from yesterday’s poll.

Truth is they work pretty much the same way. It’s just files, folders, tools and instructions - as I went over in this post.

But my own personal agent is pretty messy and often mentions stuff that’s irrelevant. So time for a fresh one.

This is how I set up my new personal agent and very lightweight memory (on purpose). Oh, and a complicated tangent!

Share Ben's Bites

Starting from scratch

Before chatting away I sketched what I thought felt like a good starting set of files it should have.

These will all go into a new folder and then I can point Claude/ChatGPT/etc to that folder and starting chatting.

• AGENTS.md - the main instructions. who I am, how I want it to work with me. questions get answers, not changes, etc. pointers to the other files.

• code.md - building preferences - I like using Vercel, Supabase, etc

• todos.md - current work

• memory.md - a pointer file to other files with memories of my stuff (about me, travel preferences, what bens bites is, what my fund stuff is, etc)

• log.md - a log of every session

The session

I created a new folder ~/bitess and started a new thread in it.

It’s recommendation was to blend my building preferences into my main instruction file, but I use my personal agent for more than just building (as we’ll see later). So I think that’s a bad suggestion.

This is probably because I’m using Codex - the ‘coding’ variant of this agent. So it literally will have in it’s system instructions “You are a coding agent”. Remember, everything it reads can guide it. So it thinks “I’m here to code, lets put code in the instructions too”.

Under that recommendation it also told me to drop log.md, because git history can record most of the work.

Woah, ben, what’s git?

git is a tool that saves versions of your files. To save a version = ‘commits’. You can look back through what changed to see the difference (the diff - literally what lines of text were changed) and go back to any point.

Agents know how to use git really well - you’ll pick up some terms to guide it.

So if changes to my files and folders get committed, the history of previous versions already exist. A log file would be essentially a duplicate. Fair. log.md not needed.

If git is added to your folder, we call that folder a ‘repository’ or ‘repo’. Don’t ask me why.

Memory

All alone in the moonlight. My initial idea was to have auto-saving memories. Every day and week with my agent would help it ‘learn’ (read: read text) more about me. It would stay relevant to my latest work, ideas and info. Brilliant.

So I asked if auto-saving (auto-commits) would be a good idea.

Trouble is, I don’t know what I want when it comes to memory. I think I’ll do a deeper exploration into memory in a dedicated post another time.

My previous folder had instructions for the agent to log things to its memory if it felt important context on me or my work.

It felt like I was getting lots of steered responses when I often want agents to help me brainstorm new things or directions - but it kept getting swayed “well this is what it says you like so sticking in this lane…” NO. Remember agents read text and that’s all they know to respond to you.

This was my idea of the log.md file - but we just figured out we’re not going to bother because it’s in the git history.

Maybe I don’t want automatic memory. Instead pick the smallest files possible with the least amount of context, make sure I know what is in those files and update it when something changes (or my agent starts saying shit I don’t like). Let’s try that instead?

Remembering chats

I asked my agent to look through a number of recent chats and see what kinds of tasks I generally do with it.

It’s hanging on to the auto-save suggestion I made and now recommending another, complicated looking, auto-commit instruction.

I often start chatting in this folder and then decide I want to build something and create a new folder for that.

This is mostly what I need most of the time:

But agents save all your chat sessions in files on your computer. They can be searched. ‘Where was I talking about that thing?’ can be answered by the agent searching past sessions. It doesn’t need saving into another file.

^{They’re often in hidden files called dot files. Which look like ~/.agents/ ~/.codex/ ~/.claude/ you can see them when you’re in Finder by pressing cmd+shift+.}

When I was asking Codex about referencing other threads it mentioned SQLite - which is a database. Essentially a spreadsheet on your computer with tables that have rows of data with your threads, id, date etc.

So I thought a database of my agent sessions could be good as an agent could search the database instead of loading files into context it can’t then forget.

I may still do this - has anyone else done it this way? Let me know!

But for now I’ll ignore the eager coder I’m chatting to telling me to do it - it can search the threads itself.

This is getting into all sorts of weeds we don’t need to be in. Its suggestions don’t always take you down the right path - agents are agreeable, as we know. Ask them what’s necessary, what are the tradeoffs, and make your own judgment.

Out of the weeds

So I stepped back - what files do we need for this agent?

Back from our curiosity detour with just files.

But I decided memory files should probably be organised nicely in a sub-folder.

The only thing in my ‘CLAUDE.md’ is the text

As Claude only reads CLAUDE.md instead of AGENTS.md (I know, stupid), we just say nope go find your instruction file in agents.md instead.

So this is the vague structure and the memory pointer file. Arguably, I could just put these instructions in my AGENTS.md

Then I got some info pre-filled, and went and manually edited them myself.

Time to save my work. Commit!

Instructions vs skills

This is the bit I’m still working out for myself.

These personal agent bots revolve around having named agents in the sidebar - Andy for accounts, Emily for emails, whatever they’re called. I think it does something to help people consider them like teammates and makes it easier for less-technical people to get to grips with it.

But they’re just chat sessions with specific instructions.

If that bot is supposed to do a thing or set of things that’s a repeatable process; label and archive emails, manage the books, etc. then they should be skills.

And skill.md is just like agents.md - it’s a file with a set of instructions!

So your main agent folder could have folders of agent folders within them, each with their own instructions (job description, even it’s own personality), and skills it can use.

But if my main agent folder is already set up for me, then having the email-skill available is enough for me to get that job done.

So why is it that people like these bot platforms like Grok Bot over their own setup they can see, edit and control. Maybe it’s just ease of use - which would be a decently good reason.

Another thing I touched on last week is that these specific task-bot agent folders (names getting silly now) have their own specific memories. So Emily is just always remembering (read: reading text) about your emails, it’s all ‘she’ thinks about! Poor ‘girl’.

And those agents can also access a general shared memory sitting higher up in the folder tree - here’s who ben is and what he does.

I think, for now, I’m going to stick with one agent, small files and folders, build up skill files for specific tasks, and think a lot more about what I’d want memory to actually do (+ where I want it to kick in to be useful).

Just to show it works in Claude Cowork and ChatGPT Work the same way:

Behind the scenes

This is how this post came together 😂

This whole format is still something I’m finding my groove with. Any (lovely, constructive) feedback is welcome in the comments.

I think it’s interesting to watch an agent session and notice things. I’ve learnt everything by paying attention, asking questions and generally getting into a mess then working my way out. I highly recommend it.

Subscribe now

• Find me on X

• Read about me and Ben’s Bites

Results from the poll are in. 996 total votes.

‘Do you use a personal agent?’

Not interested folks, r u ok? Just kidding.

Personal agent’s aren’t much more than a number of files and tools connected to any agentic tool you choose.

Share Ben's Bites

I’ll do a walkthrough of a session I had to setup a new (clean) personal agent for myself in tomorrow’s post. It really is just a bunch of files and tools. I shared my messy folder last week and I realised I must clean my desk before I start any work 😬.

Files give it context in the session - even memory is just a file it reads at the start of the session to seem-like it has memories, but it hasn’t. It’s just read those things (like me last-minute cramming for a test).

There’s so many tools out there with a bunch of features but ultimately if you have a folder with these pieces, you can have a personal agent. And that agent can use whichever tool you want - don’t like Claude’s latest writing style (pfft who does)? then point ChatGPT Work at the folder and work from there.

It also doesn’t matter if it’s the cowork or code versions of the agents.

-results in tomorrow’s post.

btw the latest best line I added in my AGENTS.md (agent instructions), from my agent spring clean:

Agents actually just answer the damn question now instead of changing a bunch of shit and then answering!

Ben’s Bites is brought to you by Name.com

Ship domain integrations in hours with the name.com API. Use the API that powers Vercel, Lovable, and Netlify’s domain services.Built for agents and developers with OpenAPI spec and MCP support.Integrate search, registration, and management. Start building.

Headlines

Two new Mac apps today:

• Meta AI - ChatGPT-like Mac app, not built for coding. It also bundles in Wispr Flow like universal dictation in the app. I tried it for a minute, but currently no reason to go back to it.

• Berd by Block - It combines the folders as project approach (codex/cc) with characters as agent (grok bot) with a board where you can pin these projects, agents, checklists and sticky notes. You can use your codex/claude code subscriptions for the tokens. Here’s a demo.

OpenAI paused improving their models for two weeks after the Hugging Face incident and signs that Astra may cross its critical-cyber threshold.

Ramp launched router.com - it picks the best model for your task automatically. They claim 40% lower cost for the same outputs. Interesting timing: Stripe bought OpenRouter the same day, and their letter to investors explaining the acquisition got leaked. Oh, and Stripe co-led Ramp’s $115M Series B in April 2021.

Anthropic’s new research put Claude on duty to help life scientists across two tasks: design protein binders from scratch and accelerate chemical analysis. Claude passed both. Reminder: Dario’s latest tweets had one clear message: “labs can’t just say AI will cure diseases, they need to show results.”

Slack is the new IDE... Slack now lets you multi-player code/build/agent from within Slack. Say Slack one more time…Anyway, it’s interesting because you (humans) can collaborate but also with agents like Claude. And when Shopify posted about their Slack agent, they said it helped so many other team members learn how to use agents. This is a pretty good direction for Slack, I think.

My feed

• All of your company’s code and context belongs in a monorepo. A monorepo is just one big folder with mini sub-folders.

• Can Grok Bot grow a newsletter to 20k subs and $20k MRR?

• Replit is giving away GPT-5.6-Luna for free (on its paid plans) in partnership with OpenAI. Keshav predicted their future…

• Vercel built its own version of Pi - a minimal coding agent harness.

• A moat is just one of 80 ways to win.

• Stop letting your agents ship ugly UIs.

• ElevenLabs’s new conversational model can help you create apps like ChatGPT’s updated voice mode.

• There’s a new AI lab for “better writing” that fools Pangram. But is that the measure of good writing? Tons of ways to fool Pangram and other AI detection tools btw…

• Notion’s solution for shared memory for agents is to create a lore based on your past conversations.

• Turn your codebase into an isometric city and explore its architecture.

• Matt Pocock’s skills have 200k+ GitHub stars. Theo walked through the select few he uses.

• A Mac app for transcribing multi-hour recordings, with a CLI for agents.

• Long read from Cursor on how they built Origin as if it were a database. Really good visually interactive post.

• Another code-storage startup nipping at GitHub’s heels. This one I’ve been eagerly awaiting, signed up yesterday and already started a paid plan. The Pierre Computer company put out awesome tools, and they’ve got a lotta taste.

• GPT-3 moment for teaching robots tasks.

• Lease a factory from Warp for your agents to build your product.

• An open-source version of Grok Bot.

• An open-source Mac app that automatically tracks your entire workday.

Subscribe now

Afters

*disclaimer - I’m an early investor 😉

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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 want to know who uses agent products as a ‘personal agent’. ie organise stuff for you, things outside of work, handling email, doing stuff on your computer, etc.

- answers revealed in Thursday’s post.

Ben’s Bites is brought to you by Cloudera

A staggering 95% of organizations have delayed AI projects due to poor governance and compliance issues. So, what needs to change? Read up on exactly that in Cloudera's newest survey, The Great AI Re-Architecture.

Headlines

Grok Bot is taking over a lot of the ex-OpenClaw crowd. Just like the last cycle, there’s already a social feed for Grok Bots that you can’t read (easily) as a human. Hermes Desktop, another work-with-agent app, also launched a Bot mode to mimic the UX of Grok mode.

Grok bot in action:

1. enterprise GTM at SpaceXAI

2. a new chief of staff

3. Marie Kondo for your inbox

4. run admin for a SaaS business

5. helping a plumbing company owner

Computer history in Codex/ChatGPT - New feature that turns your activity across apps and websites on your desktop into memories and a timeline for reference when you ask it to do something (remember the limitless app?). It’s opt-in via settings. Explainer of how it works and an example you can try.

More Codex:

1. How to enable a 1M-token context window for Sol. (I wouldn’t bother - the default is there for a reason)

2. Ultrafast GPT-5.6-Sol in the API for select customers.

New in Claude Code:

1. /design skill creates editable UI artboards where you can pick a direction and tweak it visually (just like Claude Design). Once you’re happy, Claude Code can implement it.

2. Hit your limits? Set up auto-continue for when the timer resets.

Google released Gemini 3.7 Flash, three weeks after 3.6 Flash. Big improvements on benchmarks (better than GPT-5.6 Terra and Sonnet 5). Flash 3.7 and 3.6 are 50% off till the end of the year.

Share Ben's Bites

My feed

• Now that Airtable’s been acquired, Google Sheets is dancing on its grave w/ Sheets Canvas - apps inside Google Sheets that sync with your data.

• A little Mac tool to schedule boring tasks & headless agents.

• GitHub was down (again) yesterday, and Cursor launched Origin - their competitor. Timing’s phenomenal - I half think they had it ready just for when GitHub went down, because it’s happening a lot atm.

• Theo Browne broke down his AGENTS.md and SKILLS.md files. Which reminds me, I should do this too (again).

• Factory opened up Guild, a highly selective builder program.

• A template to build your own software factory.

• Why model routing must be in the harness.

• A collection of minimal, animated chart components.

• Notes on building products people love when software gets cheap.

• Own your intelligence: a how-to guide.

• Now harnesses need a harness?

• Filling in forms should be smarter (and less annoying for us).

• Codex as an assistant video editor for the OpenAI team.

• AI usage patterns in software teams based on Linear’s data.

• Small things you can do right away to make your interfaces better.

• The ElevenLabs MCP is now available in Claude.

• Stripe has finalised an agreement to acquire OpenRouter for more than $7B.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

It’s my wedding anniversary, so while you chew on this post, I’ll be chewing through a delicious lunch in the sun.

Last week I walked through a real agent session and explained the loop as we went. Agent thinks, uses a tool, thinks again. Context fills up. Compaction wipes the session. If it’s important, put it in a file.

Today’s post is why the file bit is useful when thinking about what a ‘personal agent’ is.

You may have heard about OpenClaw, Hermes, Grok Bot (launched this week), and wondered why they had/have? such hype. Why did/do? people love them? Why are they different to using ChatGPT work/Codex or Claude Cowork/Code (god someone do something about these names!)?

Do I need a personal agent too?

They’re not different products, per se, they’re just packaged like one. It’s really just about the setup. You can set up the same thing, it’s just a file and folder system.

Share Ben's Bites

So what’s in a personal agent?

It’s what most agents have:

• instructions

• tools

• context

There’s nothing really special or unique to being a personal bot.

What it looks like

There’s two paths for a personal agent.

One main agent, like ‘Jarvis’:

Or split them into several agents with their own files, personality, job descriptions - just by editing the instruction file.

People often split agents to be task specific. e.g. You are my money manager agent, You are my marketing agent, You are my copywriting agent.

You can give them job descriptions, names and personalities by customising their instruction file.

I use one agent generally as I like the ‘Jarvis’ approach. I’m a poor delegator. If I need task-specific work done the main agent just reads a task specific file e.g. newsletter-research.md.

But I see the value in splitting them as ‘separate agents’, each with their own thread. This is how Grok Bot works - each agent has its own thread, but you could just set up pinned threads in Codex/Claude pointed at each specific agent folder you create.

It’s all about folders and instructions.

An agent pointed at a folder reads its instructions and that’s all the context it has.

File and folder organisation is helpful to stay on top of, I’ve not been great at that so here’s what my actual ‘chief of staff’ folder actually looks like:

Memory

It isn’t really memory, it’s a text file with a log of things that happened.

The agent reads it to get up to speed to ‘remember’ what were the recent things you worked on, who you are, etc.

Personal agents like Grok Bot give you several agents which all have their own agent-specific memory, i.e. money agent remembers all the money stuff.

But they can ‘share’ memory too - which is just agents reading other memory files.

If the personal-agent instructions says to save important information to the shared memory, each nested agent can decide when information in a session is important enough to save there.

Tools

Codex/Claude Code have a computer use tool where it can actually navigate your computer like you (but better?).

Grok Bot uses a shared computer. All the agent chats run on one machine: same files, same installs, same logins. Each chat gets its own screen.

Computer use lets it do the thing a person could do. Navigate websites, install apps, fill in forms, create files, etc.

Routines and automations

You can give all of these agents automated tasks to do, workflows, scheduled, recurring, whatever. You just ask them for it.

They all have the capability to do so given they have tools and can use a computer.

Some automations start a fresh session with the same instructions. Others wake an existing thread and continue its existing work.

You can just ask your agents to set up automations for you or Grok Bot has a ‘teach’ feature where you walk through a task on your computer recording yourself and it’ll learn how to do it. You could just record yourself and give to any agent to analyse and turn into a workflow. So it’s not a new capability, but it feels easier to try.

Wrap up

So you can see they all have very similar capabilities. They read files, take on a personality and job if it’s in its instructions, write down memories and get to work.

They can send messages to other threads or agents to get info or delegate work.

It’s all really a setup of folders and files.

I prefer having more control over the setup - namely, what goes into each file, but also changing the model or reasoning for certain tasks (which Grok Bot doesn’t have) and thinking - I like to see what my agents are doing.

And agents can be used for all sorts of things, again its just the setup. Does it have the instructions and tools available? If it does, they can do pretty much anything you’d want them to do:

• Find/book flights

• Order food

• Negotiate contractor quotes

• Organise files/folders

• Collect invoices from emails and file away

• Create decks

• Create a morning brief

• Build websites and apps

The set up is basically:

• Create a folder for the agent.

• Add an instruction file that defines its job and rules.

• Add a user file and a memory file.

• Open that folder in Codex or Claude.

• Pin the thread.

• Give it tools and permissions.

• Add a schedule when the task is repeatable.

Behind the scenes

This is how this post came together 😂

Subscribe now

• Find me on X

• Read about me and Ben’s Bites

I had early access to Grok Bot, like many others on X posted. Except, I didn’t actually use it 😬. But since it was publicly launched I have started using it…and, I like it?!

There was a wave of ‘personal agents’ with OpenClaw, then Hermes (not the luxury brand) with many flocking to them. I used them to begin with but then just stopped completely. I can’t quite put my finger on why - one may be because I was just generating shit for the sake of it. Having a shower? Make sure my agent was doing something. Going to sleep? Get my agent on something. Things I used? Zero.

I’m a one-session-at-a-time kinda guy. I hate having to review a ton of agent work (or any work)!

Codex/ChatGPT with its mobile connection was all I needed from an ‘always accessible’ agent that could do stuff on my computer. But Grok Bot’s form factor is slightly different, and it’s pretty nice.

You ‘personify’ your agents. Each gets a system prompt e.g. “you are ben’s money manager agent”, they can send messages to one another, they have access to their own virtual computer (which you can watch them use), you connect your accounts like Gmail, Calendar, Slack, etc (though I ran into several hiccups trying to connect accounts).

I think it’s just less is more. There’s a lot less complexity with it, it feels like you’re messaging a teammate in Slack. You can set up automations really easily with it, and you can even go into ‘teach’ the bot by it watching your screen.

You can effectively do all the same stuff in Codex or Claude, but it’s just a couple of steps more complicated. Maybe I’ll try and mimic the setup in Codex?

Currently, you can only access it if you’re on the $200 plan for Cursor or Grok. And of course Grok Bot is powered by the newly released Grok 4.6… (with 4.7 a couple weeks away). More on that below.

Ben’s Bites is brought to you by Name.com

Ship domain integrations in hours with the name.com API. Use the API that powers Vercel, Lovable, and Netlify’s domain services. Built for agents and developers with OpenAPI spec and MCP support. Integrate search, registration, and management. Start building.

Headlines

Grok 4.6 is a damn good model. It’s near Sol and Fable’s performance on benchmarks, and it’s way cheaper than both. Also see: Grok 4.6 – A field guide.

SpaceX + Cursor is now the #3 frontier lab besides OpenAI and Anthropic (rip Gemini, unless their next model is actually good).

Elon says Grok 4.7 will be ready in the next 3-4 weeks, and it’s currently being post-trained on SpaceX’s company data to make the model best at real-world engineering, if not everything. The model card for Grok 4.6 also covers many benchmarks beyond coding, where it’s either #1 or #2.

ChatGPT now lets you import and sync your projects, agent sessions, skills and more from other products like Claude Code.

Your chats started via Claude’s Chrome extension are now saved in your account and can be accessed from your Desktop, mobile or web app. That also means the Chrome extension chats can do the work you’d usually open Claude Cowrok for.

Share Ben's Bites

My Feed

• An open source version of Grok Bot.

• /interface-review - get a detailed analysis of your UI, typography, layout, colour, writing and accessibility.

• /show-me - read less text by making your agent respond in these compact styles to present information - component/file trees, diagrams, pseudocode, and more.

• Ref - shared planning mode for your team before agents start to build the thing. (read more)

• treg - 2,600 agent-friendly tools for seo/geo, social, leads, ads and more.

• Caffeine - Keep your Mac awake while you work, then let it sleep when your coding agent waits for you.

• Ori Pi - Run Pi directly on all models available via OpenRouter.

• Natter - type anywhere on your Mac using your voice, with live transcription. (how it’s different)

• Ramp AI Index: disappointing adoption of Fable 5.

• givemeanode - rent GPUs & CPUs with the infra handled for you and then let your users use them without any worry.

• Hone - AI that works towards outcomes that help your company.

• Claude fixed my WiFi from 200 to 810 Mbps.

• Vercel is using a software factory of agents to build their AI SDK, with 35% of PRs coming from the factory.

• Beating Pangram’s AI detection by training a model over the weekend.

• How to keep thinking.

• Everything hackable will get hacked.

• A complete guide to AGENTS.md.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

AI outputs are becoming harder and harder to read. I don’t just mean their sloppy smell but a lot of the time I just want to scream “speak to me like a normal human”, which, I guess, is ironic.

So I’ve been testing two new instructions. The output’s been 10x better for me, especially when it’s something “I’ve” built. If I say “I’m not technical,” it uses analogies for everything, and honestly I think that’s worse than the slop.

This is what it’s like when describing a Chrome extension I built (which I discussed in Friday’s post*).

You just add this line to your custom instructions (in settings) or your AGENTS.md (it syncs both anyway).

wtf is ‘ASD-STE100 Simplified Technical English’? It’s a writing standard. It makes technical information easier to read, understand, and translate.

I also added:

The two instructions in combination are the sweet spot for me. The ADHD line gives more succinct bullets and headers. It’s a skimmers dream!

*I didn’t realise all the images were transparent so when you enlarged them they looked like shit. I won’t do that again.

Meanwhile my wife thinks there ‘must be someone or a robot’ at the other end of her chats.

Ben’s Bites is brought to you by Cloudera

72% of enterprises say their current data architecture needs a significant overhaul to meet AI goals. See how that shift is taking shape in Cloudera's Great AI Re-Architecture report.

Headlines

OpenAI updated the availability of its models in the good ‘normal’ ChatGPT. Paid users get a special version of GPT-5.6-Sol as the default, and free users get unlimited chats with GPT-5.6-Luna (likely on lower reasoning levels).

Plus, they’ve expanded their cybersecurity program Daybreak and released GPT-5.6-cyber with it. Astra, OpenAI’s next mainstream model, is critical for cybersecurity, so we will have to wait a bit for it to release.

Bloomberg released a scoop on OpenAI’s secret new device - it’s likely a smart speaker without a display in a doughnut-like shape. Easy to carry, over $300 a piece and coming in 2027. (non-paywalled version)

Meta is back to open-sourcing models - Muse Glimmer is their new open-weight model. 30B parameters and tuned for agentic tasks (including computer use). Meta says it’ll soon release the weights of Muse Spark 1.2 - their best-performing model till now. Zuck also wrote a 6000-word piece on his and Meta’s philosophy and values on AI. TLDR: Mark thinks everyone will have

• personal agent that understands them and their goals

• tools for creativity, business and education

• free or affordable access to these tools

Grok has a new image generation model - Imagine 2.0. It is #3 on the image generation leaderboard and #2 on the image editing leaderboard managed by arena.ai.

Your sessions in Claude Code can now message each other.

More Claude news:

• Anthropic will watermark all Claude-generated text (🤮)

• It is planning an IPO in September/early October (I hope my SPV shares count!)

• One of their unreleased Claude models did some more cool maths

Share Ben's Bites

My Feed

• Reliable AI is hard. Nutrient converts PDFs to JSON or Markdown, with coordinates + confidence per value. Grab an API key & try it for free.*

• Raindrop Signals 2.0 - uses custom classifiers to find problems in AI-agent traces with near-frontier accuracy at much lower cost.

• Every company needs a Cassandra - short read, worth your time.

• Agent Plugins - standard for packaging MCP servers and Skills. Supported by Codex, Cursor, Copilot and more.

• Stagehand - the SDK for browser agents.

• Claude Code for normal people with Grace Clark.

• Energy - Another desktop app to work with agents. It creates job-specific assistants instead of using your folders as the home for projects.

• Gloss - prompt any website into looking however you want. Installs as a Chrome extension. (see example)

• seoskill.dev - The only SEO skill your agent needs.

• Replay - Mac app to download videos from YouTube and X to watch later.

• Mintlify Index - search layer for technical knowledge from 200k+ orgs. They claim 2x faster results and 9% higher accuracy vs Context7.

• Lindy Teammate - Let anyone from your team text lindy from Slack like they would another employee.

• The age of “no code” has passed. Or maybe you can build a new version of Zapier. (don’t though, I’d like my shares to be worth something)

• Kitesurf by Cloudflare - a lightweight web browser for AI agents.

• Spotify has a coding app now, Xirp - it combines all your work across Claude Code, Codex, and Gemini CLI in one place.

• Designing with AI? Make a jig, i.e., a tool that makes it easier to make something else.

• AI adoption is a myth.

• Friction as a feature.

• Siri sucks. So I made Codex act as my iPhone’s voice assistant. Here’s how.

• We’re not using enough inference.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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’m trying something new - this email walks through one of my actual agent sessions and I’ll explain what’s happening along the way. The build or task I’m doing isn’t important. But I’m looking at how I could be using agents more effectively.

You might pick up a thing or two, I’m finding it helpful to solidify my own learning.

Please comment what you thought of this, was it helpful? anything unclear? want to see other things?

Subscribe now

What I was trying to do

I was setting up a bookable appointment link in Google Calendar and got annoyed that you can’t just drag time slots on the calendar grid, you have to type each date and time range into a clunky form. I wanted something that would let me drag slots directly on the week view and have the form update automatically.

So I fired up ChatGPT (I use Codex mode but works in ‘Work’ or Claude Cowork etc). I wanted to test Luna on Max reasoning as the price has been cut 80% and people have been saying how great it is to use.

It’s not the best prompt, I’ll admit.

But it gives the agent enough understanding of what I want so it can explore options. Plus a screenshot so it knew what screen I was on about.

This kicks off the ‘agent loop’. The agent thinks about what to do (what can be done with Google Calendar), then acts by using a tool (in this instance, web search) to gather context on how to solve my task.

The websites it read are now in the context window. I didn’t look at them so I have no idea what info it found or if its true. All the text it read is now in its ‘memory’.

Imagine 20 websites went in, there could be wrong or contradictory info that could mislead the agent. This is why you hear so much talk about context. It’s important, and you want it to be full of the best possible information.

Agents often do many loops for a task. They’re ‘go-getters’ by nature. Which is why they need babysitting.

They keep looping, gathering more context until they have what they need to complete the task. For my fantastic prompt, 55 seconds and 2 web searches was enough.

Build it

The agent came back with a mini plan. I skimmed it, as usual (which cost me...).

I overlooked one point which was not how I wanted the extension to work, it should create the times automatically in the form as you drag tiles, not manually click to sync.

What I should’ve done is gone back and forth to ask how things would work, maybe mockup some wireframes I could annotate with feedback.

But I didn’t.

I just said build it...

It cycled through it’s loops and it was built!

Ha, not quite.

My first thought here was:

1. I shouldn’t need to install this myself

2. If its not installed, the agent can’t have tested this live

3. Why did I just say ‘build it’!?

The agent has tools it could’ve used when looping over the task, specifically Computer use and Browser use. It could’ve installed it and tested it live on my actual calendar page.

It didn’t, so I knew there’d be hiccups.

I installed it and tried dragging time frames but the form syncing didn’t work (shock).

What I should’ve done is say something like

That would’ve saved me time and tokens...

I went through my frustration escalation.

• I start by typing the issues I run into.

• 14 minutes later still had issues. I moved to stage 2 - voice ramble and a screenshot.

• 61 (!!) minutes later still had issues. I moved to stage 3 - I record my screen with a voiceover, pointing my cursor at moments with issues. Agents can break videos down frame by frame and transcribe to pinpoint what you’re talking about.

Each turn (back-and-forth) the context gets fuller. After the first set of issues I sent it, it started actually using Chrome to test, thankfully.

But from reading it’s thinking, it didn’t test fully as it didn’t want to override my ‘work’ but that page was for the agent to absolutely use, that’s kind of key to test if this thing worked properly.

So on the second attempt at fixes I rambled a voice note and added a screenshot. This time I added things it should check and tests it should do.

This is that verification layer an agent considers when it’s thinking about the task being complete. Do all the tests (that the agent comes up with) pass with no issues = extension works.

When giving an agent a task, you should think about what criteria would mean this task is ‘done’. For a website it could be that all the content is formatted well with spacing, your design system, and works on mobile. For email triaging it could be that all the emails in your inbox have a label and are moved to the correct folder.

Verification is something I’m still working on as a lot of my tasks are not code.

Development tasks are easier in this respect - agents check ‘did this code run successfully’ (ie. does the code work).

You can also ask your agent to help come up with the verification criteria…

The agent was looping for an hour (!!)... I should’ve known something wasn’t right, but I thought maybe it’s just being really thorough 😅.

It turns out the context window hit the limit 3 times in that turn.

When the context window gets full (~250k tokens), it summarises everything, but keeps some things intact like your messages and the system prompt (the instructions the agent has baked in already).

In this hour, the agent had been testing a lot and finding out things to help it test the extension properly.

But then compaction wiped it from the context so it had to relearn it again, and again.

A bigger model may have helped here as it could more easily recognise what information would be helpful to keep through a compaction. The agent harnesses themselves all compact context differently. Most of the time I don’t worry about it, but this was an example where I should’ve said ‘when testing, note down your approach and learnings’ or something which the next context could’ve picked up.

If it’s important and in your context window, put it in a file! Files don’t get wiped in compaction, only what’s in your session.

I’d also noticed in the agents thinking that it had spent a bunch of time (and tokens) testing an older version of the Chrome extension, it had been saving each new version as a separate folder. Wtf.

This is where I switched models to Sol with High reasoning, Luna Max is fine for everyday tasks, searching, and things like that, but for code and testing it didn’t work well for me here. Skill issue?

I hit level 3 of the frustration escalation so I added a screen recording.

And just to clear up the older versions I followed up with a message mid-task.

Most agents have this in settings. I often send a prompt and then think I better quickly follow up with some extra info or another ask. This sends your prompt whilst the agent is already working, so you can steer it.

13 minutes later it was done. It worked exactly how I wanted it to. Good boy Sol!

So the extension works, it’s useful. I may use it twice a month. But my process wasn’t great - a lot of reminders in there for future sessions.

Bugs are inevitable, for everyone. You just have to think about how you can provide the best context for the agent to figure out its way to fixing them.

What I should’ve done:

1. Read the agent’s mini-plan before saying “build it”

2. Clarify the details upfront instead of discovering it as a bug

3. Tell the agent to install, test, and iterate in Chrome from the start

4. Use the bigger model for iterative debugging loops

5. Be explicit about compaction: tell the agent to preserve key learnings

6. Check the agent is editing and running the same files

None of these are about the agent being bad. The agent was actually really capable — it built a working Chrome extension from a screenshot and a two-word prompt, it controlled Chrome to test its own work, it diagnosed bugs from a screen recording.

The inefficiency was in how I steered it. Every extra hour and every wasted token came from a decision I made (or didn’t make) about what to tell the agent and when.

Here’s the demo video it produced:

Did you like this type of post? Comment and let me know 😊

Behind the scenes

This is how this post came together 😂

Share Ben's Bites

• Find me on X

• Read about me and Ben’s Bites

This time last week I told you I’d downloaded t3 as my ‘all-in-one’ agent app, which lets you use your chat/claude subscriptions from one place.

Well forget that. (strong opinions, weekly held)

There’s a new kid on the block and I’m in love…with the name, bb, but also the app itself.

If you use claude/chatgpt/pi/cursor/factory/[any agent] - you can use it within bb. It looks and works a lot like the chat/codex app (great). I got it set up on my mobile in like 4 seconds which is miles above other apps I tried (t3 was finicky for me here).

But the biggest thing of this app is its extendable nature. There’s a lot of talk on extensible software right now. Pi is one of the best coding harnesses out there and it’s build to be fully customisable (btw is it extendable, extensible, are they the same?!) - the agent can improve its own tool.

• The minimalism of Pi makes it more performant on tasks vs other harnesses.

So much so that companies like Prime Intellect built their agent on top of Pi, as well as many other companies.

• Prime Agent - self-improving harness for coding and long-running tasks for RLMs.

BB is built the same way. It’s an ‘IDE’, except that it isn’t, IDEs are older tools like VSCode that I wouldn’t want to be caught dead in these days. The ‘IDE’ of today is these desktop agent apps that we’re familiar with. BB can extend itself, you want a plugin? Ask it to build one (I got prime-agent to build a factory plugin for it). Want a task tracker? It’ll build one.

I’m never going to spend 100% of my day in an app where I cannot switch models (and agent harnesses), I just cant. I like GPT models for some things, Claude for others. Sue me.

I think the software that we’re going to want (ie builders but not developers) looks like mini ‘widgets’; todo lists, emails, feed readers, etc (like my canvas exploration from last week - which I’m still using btw) and bigger tools like my twitter bookmarks and trends tool (which is private but I may open up one day).

And the tool that you spend most of your ‘work’ time in should have the ability to support adding extra capabilities to it as you learn what you want or need to get your shit done.

Semi-related, there was a few posts on X about why agents aren’t being used outside of developers and early adopters. A quote on the original said:

“You cannot make the masses want an automation flow when they want to hang out w/ their friends and not think about their job.”

I 100% agree. And I still think the pool of ‘builders’ will skyrocket. My definition of that is loose - anyone who wants to use agents to build tools for themselves and/or others, for fun, work, life, whatever! Where do you land in that?

In fact, I’ve believed in this pool of builders for about 10 years now but no-code was too clunky and limited to unlock them, AI is so much easier. It benefits from word-of-mouth. You don’t have to ‘learn’ anything per se, but you start to feel a way of working with it and pick up stuff along the way. Like learning a language, if you’re in it you’ll pick it up faster.

Ben’s Bites is brought to you by Type

Work is multi-player. Now AI is too.

Type puts your team's best AI work into a shared space. Bring your company's knowledge, integrations, and skills together so everyone can use them with any model. Try it now.

Headlines

Bending Spoons is acquiring Airtable, for far less than its last valuation. Though Hyperagent, Airtable’s new AI business, was spun out and is not part of this acquisition. More details on the deal. Another casualty in AI’s wake.

To jump on the news: Obsidian created a free tool to export your Airtable content into Markdown. Another one: Cloudtable - An Airtable clone but it runs on Cloudflare (repo).

Google’s Chief Scientist, Jeff Dean, is leaving the company after 27 years to start a lab of his own with three colleagues. The new company, Discovery Loop, wants to automate machine-learning research. This means a lot of internal moves inside Google. Most notably, Demis Hassabis is stepping away from GDM’s day-to-day leadership to focus on long-term research.

Cloudflare has been releasing a ton of agent-first tools recently: Cloudflare OS is a shared AI workspace with connectors, plus Agent Tracing lets you replay agent sessions and follow them from invocation through model calls. Cloudflare Wallet is another initiative from them to give agents identities, spending limits and wallets for paying for services.

Gemini Notebook (i.e. NotebookLM) now lets you create the different types of artifacts it supports (quizzes, slide decks, etc.) right from the chat window, instead of the side panel.

Meta launched Muse Code, a CLI coding agent powered by Muse Spark 1.2 - a new model with coding-related upgrades, bringing it close to Grok 4.5 level of performance.

Goodfire AI made their interpretability and training tool, Silico, generally available. It’s $1000/mo for individual researchers. These guys tried Silico partly while teaching a 30B model to play a board game, worth reading their experience.

Subscribe now

My Feed

• Capsule: Video Skills and agents for enterprise teams who don't want to sacrifice brand quality.*

• Atlas by WorkOS - an AI teammate in Slack that answers questions, takes actions and learns how your company works.

• Context.dev - API that turns the live web into structured, continuously refreshed data for agents.

• Notion’s results from a survey on 6,000 professionals across 10 markets on how AI is being used at work.

• givemeanode by SF Compute - give Claude, Codex or any agent H100s for training, benchmarks and batch jobs, billed by the minute.

• Power 2026 - primer on power pricing and data centres for understanding AI’s electricity bottleneck.

• Not Diamond Code - model router for coding agents, with claims of 20-65% lower cost on long-horizon tasks.

• Recursive self-improvement is limited by verification, not computation.

• Running Streaks - Opus 5 turned runners’ data into an interactive D3 visualisation using this prompt.

• NYC AI Atlas - open-source 3D map of New York’s leading AI startups.

• Two more cybersec incidents at OpenAI came up in external cyber evaluations.

• Mistral launched Shieldstral - a small 3B open-weight model for content-safety. Easily runs on-device, even on a basic laptop.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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 wasn’t prepared for AI hard-hitting truths this morning but I tested this:

In the link, you download the file ‘reflection engine’ which is just a very big prompt basically. Upload the file to your agent and ask:

Please evaluate the attached markdown file and complete all tasks.

I tested it with Fable High and Sol Max, and Sol produced the better report. It was more coherent (agents are speaking more gobbledy goop these days) and nailed connections. Fable’s was harder to read.

I’m going to share some of my report here, because we’re friends - we read and we don’t judge.

So yeah, there was a lot in here and now I’m having a 40+ question ‘grill-me’ session with my agent to address it all and come up with a plan 😅.

I actually found it really telling and helpful. I do have all my therapy transcripts and all sorts of ‘memories’ (read: text in files) that it combed through to produce this.

I recommend you give it a whirl.

Ben’s Bites is brought to you by Nutrient

AI hallucinating? Your document parser is probably dropping fields.

With Nutrient, send any document and get back a JSON schema. Every field returns a bounding box, a confidence score, and a match label, including an honest not_found when values can’t be grounded. 5,000 credits a month, try for free.

Headlines

OpenAI reduced the price of GPT-5.6 Luna by 80% and Terra by 20%. Luna at max thinking effort now scores roughly the same as GPT-5.4 xhigh, the best possible model about four months ago, but at only 8% the cost.

That’s 10-12x more work (or play) you can do if you stick to 5.6 Luna with a little bit of Terra and Sol usage here and there.

In a few tasks I used Luna Max and found its great if it’s not building-related work ie chatting, research, reading/writing files for day to day stuff but I had to tap in Sol High to clean up some mess Luna made on a little chrome extension I made yesterday.

In the meantime, OpenAI also teased their new major model - Astra. They released solutions for 10 long-standing problems in maths and theoretical computer science, all found by Astra. And yes, it’s maths, not math.

Google launched Gemini Robotics 2, one AI system designed to work across everything from robot arms to full humanoids.

It has one model that turns vision and instructions into movement, another for planning complex tasks, and an on-device version. Google says it can control a humanoid’s whole body, handle delicate tasks and coordinate multiple robots.

More models to pay attention to:

• DeepSeek V4 Flash costs $0.14/$0.28 per million input/output tokens—less than Luna’s $0.20/$1.20—with a 1M-token context window.

• Qwen3.8-Max is a 2.4T model that Qwen says comes close to frontier performance for coding and professional work at $2/$6. Its weights, plus those of a smaller 27B model, are coming next week.

• Inkling-Small is open-weight already. It activates 12B of its 276B parameters, and Thinking Machines says it matches or beats the roughly four-times-larger Inkling on many benchmarks.

• MiniMax H3 - a new video model that Krea says matches Seedance’s features for less; open weights are promised soon.

• P-Image-Ideogram - four image-generation quality modes, native 1K/2K output and prices starting at $0.003 per image.

Subscribe now

My Feed

• Ben, who worked with us a couple of years ago & is now at Vercel, built this marketing team template for agents - bring work to a team lead and it delegates the work to a team of marketing agents built with Eve (Vercel’s agent framework).

• YC open-sourced a customisable multi-agent harness it uses across accounting, legal, events and engineering. Here’s the repo. Vercel also built an internal agent called “@v” for finance, comms, docs, marketing, engineering and business analytics.

• ChatGPT desktop app has a new sidebar view - press the bell-like icon, and you can view all conversations that need your attention quickly. Although I don’t understand it at all… I thought it would sort based on recency but it’s not?

  

• GenOffice - free, open-source AI office suite for Mac and PC, built it in a week using $10,000 of tokens.

• The next five years - Patrick Collison’s short survey on how AI will change wages, productivity and the wider economy.

• Conductor Cloud - multiplayer cloud workspaces for Claude Code, Codex, Cursor and OpenCode that keep running after you shut the laptop.

• The human side of AI - taste, fried attention, frozen competence and why more automation can create more human work.

• Cloudflare Computer - open-source virtual filesystem and execution runtime for agents, with container, shell and JavaScript backends on Workers.

• Devtools must be open source.

• The harness is in the app - build the agent loop into the product, instead of adding it from outside.

• The session you cannot take with you - a five-part test for whether an agent session is actually yours: inspect, export, replay, audit and delete.

• How Monologue rebuilt its site with Codex within a week.

• Comp AI’s CRM - open-source, agent-first CRM with durable research agents; MIT licensed.

• FutureSearch - ask for forecasts, including “if I do X, will Y happen?”; its superhuman accuracy claims come from the company. (launch)

• remote-agent-browser - browser-control API for agents with managed cloud sandboxes and no local Chromium install.

Afters

You should go and buy my buddy Jack’s new book - he’s a phenomenal writer and thinker. His angle on anything is always something I look forward to reading. (It’s not only for your twenties…)

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

Following on from Tuesday’s messing around building, I built a few more widgets on my canvas. It pulls my X bookmarks, emails, and todos.

Before making this video I didn’t have Loom installed (since it’s crap after being acquired), so just told Codex to build me one.

I drew the image on the left, then just sent that prompt and it worked straight away.

So as software and mini tools are getting easier to create, the tools to create them are not...

I use Codex as my default app because it’s better than all the others and works the best on mobile. But I just downloaded t3 because they basically copied the interface and features, but it lets you choose between different agents; codex, claude, cursor (pi + droid soon). It’s made by a reputable developer that you may have seen mentioned here before, so I trust it’s built well.

At the moment, I ask Codex/ChatGPT to be the orchestrator and to go ask claude about something that is design-related. Which is fine-ish but not great as user experiences go.

What I’m noticing by doing this is how well Codex creates prompts for other agents to follow. I just have a normal chat in my session, then say ‘Use X agent to implement this. Watch its progress and send screenshots every time new work has landed.’ To which it sets up its own monitoring every 5 mins or so and updates me in the same thread, with screenshots.

I’m going to try and put together a ‘bites of the week’ email over the next few days to try and summarise all the stuff going on, what themes people are talking about (loops?!, software factories?!, etc) and explain them.

Let me know if there’s anything specific you need to wrap your head around (I may need to too).

Ben’s Bites is brought to you by Brief

Struggling with fragmented context, slow product decisions, and rework? Brief distills your critical product context into an opinionated graph, then puts a PM agent everywhere you work, (e.g. Slack, Claude Code, email) reducing alignment tax and accelerating cycles. Learn more.

Headlines

OpenAI used Sol to optimise Sol itself, cutting serving costs by 20% and making it 15%+ more efficient at generating tokens. And turns out, it also tops the ARC-AGI-3 benchmark.

Well, there’s a catch: OpenAI says the official ARC-AGI harness hurts Sol’s performance by “forgetting” its reasoning every turn and disabling compaction. Fixing these two things triples Sol’s score from 13.3% to 38.3%, with 6x fewer output tokens.

Re: last week’s fiasco of an OpenAI model hacking Hugging Face - HF published a full replay of roughly 17,600 actions taken by the model. METR and Redwood Research will also independently review what happened.

Though OpenAI is not out of trouble just yet, a Reuters report claims that the same model broke into a customer account at another company (Modal Labs), with rumours suggesting that even more companies were affected.

Anthropic also claimed that Claude Mythos found better attacks on two cryptographic algorithms, though neither affects systems in use today.

Separately (not at all as a reaction to this general trend, right?), ~1300 people working at leading AI companies (OpenAI, Anthropic & others) want the US government to help “pace the frontier” of AI development. Kinda expected when the pace picks up, but this time a lot of the “model makers” themselves are in favour of this pause/slowdown.

btw, The Information reports ChatGPT is nearing one billion weekly users - a milestone OpenAI originally hoped to hit seven months ago. More from OpenAI this week: Codex Security CLI, free frontier access for Academic Researchers, and two new transcription models.

Grok app builder - Grok has a vibe coding interface inside its app now. Create games and apps that can be shared directly to the X timeline. Also see: Drawesome - a zero-dependency drawing toolbar for React, built over a weekend with Grok Build.

Pangram 4 claims it catches 98.83% of humanised AI text with one false positive per ~24,000 docs. An early test found all 38 AI-written words inside a 1,198-word story, though not on every run. Its new image detector claims 99.5% accuracy too.

Subscribe now

Quick links

• Tavus - Build AI that comes to life: video agents that see, hear, and answer in real time and do anything you want. Use TAVUS50 for 50% off.

• 66% of July traffic on docs built with Mintlify was from agents.

• Resend added an MD version of their pricing page to avoid confusing agents.

• 0%, 50% or 200% - ignore AI, halve staff or double the ambition.

• Slackbot can now run code in the background for data analysis, slide creation, and to make live reports or widgets.

• Gemini’s macOS app got a voice mode that lets you ramble, and the app turns it into a clean prompt. Hold Fn to try it.

• The AI future is for everyone - Mark Zuckerberg

• Replit Design - make sites, prototypes and graphics from prompts, URLs, Figma files or screenshots.

• What’s gone wrong with AI & labor.

• Kami - open-source Hermes agents that find customers, prepare outreach and content, then act after your approval.

• Coast - fully local memory for you and your agents, built from what you see on your Mac.

• Pragmatic leverage in the software factory.

• Crew Studio - find useful ideas where agents can help your business, build those agents with the option to take the code home to run anywhere.

• HeyGen Video Podcast - turn a doc, link or idea into a two-host video with scenes, camera cuts and B-roll.

• Copper - local scratchpad for saving answers, links and follow-up prompts across your AI apps.

• FT Chart Doctor - visual vocabulary and examples for choosing a chart that fits the relationship you need to show.

• Mitchell Hashimoto (Ghostty) and Andrew Ng (deeplearning.ai) are both starting new companies: Superlogical and LearnVector.

• MCP’s biggest update removes the need for servers to remember every ongoing connection, making them easier to run and scale.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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’m back from holiday and straight back to messing around building. I’ve been kinda obsessed and enamoured with the demos from tldraw. It’s a free canvas/whiteboard app but their offline app (recently updated) is cool because agents can do all sorts with it; create visualisations to explain stuff, make games, interactive blog posts and a ton of other stuff.

Like turning our logo into a little animating mascot. He’s called ‘bites’ btw.

or little widgets like a timezone checker

or a daily weather timeline

I’m not good using agents when everything is just text in a file. It’s boring, I don’t read it and I sure as shit don’t do enough (anything?) with it. So if it’s visual and all on one canvas, maybe I will?

My idea here is to create a little ‘one-stop’ canvas for some widgets I use regularly as well as showing me my todos/emails to sort and things like that. I’m then planning on making my personal site an interactive canvas too, then some more explainer-like stuff too.

And maybe I’ll get the mascot hooked up to an agent so i can drag him to my todos and ask him to do stuff…

I’ll share more soon 😊

Ben’s Bites is brought to you by Ability AI

This startup runs on 17 AI agents - and open-sourced the platform. Marketing, client health, competitive intel, finance - real ops, not demos, with every agent action git-tracked and a human approving anything external. Agents you own, not rent. Apache 2.0: github.com/Abilityai/trinity

Headlines

Claude Opus 5 is out - Anthropic says it gets close to Fable 5 at half the price. Claire hated using Opus 5, but when asked to rank model outputs in a blind test, she ended up rating it first.

Every’s review lists some behaviours that might explain this: it argues, stops early and fights the skills people built for older models.

And maybe we are overprompting it.

Anthropic says it removed more than 80% of Claude Code’s system prompt for Opus 5 and Fable 5 with no measurable coding-eval loss.

ChatGPT Voice can now run the ChatGPT desktop app. You can use it in ChatGPT Work and Codex both; just start a “Voice” session and speak to it. It can create new sessions for individual tasks and then report the results to you in the main thread. It feels a lot like OpenClaw with voice.

I didn’t have a great time using ChatGPT Voice for “serious” work. The delegation of tasks to new sessions via ChatGPT Voice doesn’t feel as smart as doing it in a chat with 5.6 Sol in Codex. Maybe it’s the lack of control on what models/thinking level is used for each new session.

I imagine Voice to work great for productivity type tasks, checking emails, creating draft replies, pull some info from a dashboard and create a chart etc.

For serious work, I still prefer using a speech to text app (currently using one I built — optionafk.com) and sending that to the agent of my choice with full control over model/thinking effort.

— Keshav

Also, I wouldn’t blame you if you missed it, but Claude also has a voice mode. Till now, it could only use Haiku, but now it can use Sonnet and Opus as well as call multiple tools like Gmail, Calendar and Slack mid-conversation. I don’t remember when I last used Claude’s chat option. It’s been Claude Code for a long time now.

Kimi K3’s weights and technical report are out. It’s already available in Droid at 50% off until 10 August.

Following rumours that the US might ban companies from using Chinese open-weight models, NVIDIA’s CEO Jensen Huang made his first-ever post on X - a long statement backing open-source AI. Over the weekend, everyone signed that statement (OpenAI, Google, Microsoft and 150+ more) - except Anthropic.

Of course, that led to some backlash, and Anthropic did what it does: published an essay to save face, saying it never pushed for an open-weights ban. The reaction to it is people calling Anthropic out on their hypocrisy — they are okay with open models as long as they are cute little 8b toys, but not if they are “big, dangerous models” (to Anthropic’s business).

Subscribe now

Quick links

• FLUX 3 - one model for image, video, audio and action prediction; video is in early access.

• Health in ChatGPT - US users can optionally connect Apple Health and supported medical records, then ask about changes in context.

• BackSearch - search the web as it was on a particular date in the past. Right now, it only works for news in 2026.

• here.now Workspaces - private team home for the sites, docs and dashboards your agents publish.

• Why software factories fail - part 1, part 2, part 3.

• open-git - a code host built for agents, with an autopilot that fixes your PR when CI goes red.

• OpenWorker - open-source Mac agent that finishes everyday work and asks before consequential actions.

• Pilot Protocol - a network of over 250,000 agents that you can use to gain distribution for your app. (I’m an investor)

• Notion as code - define your whole workspace in TypeScript and deploy it through the API. (beta)

• The Stack v3 - open code dataset with about 5T filtered tokens across 770 languages.

• box - full VM sandboxes for your agents at $0.00001 a second, with snapshots, SSH and a desktop.

• Mintlify - turns resolved support tickets into proposed documentation fixes.

• Cognition acquired Interaction - the team behind Poke joins the makers of Devin.

Skills section

• Interface Craft - four new lessons on advanced animation, plus an updated DialKit skill for timelines. (tweet)

• /prototype - asks your agent for several UI versions and saves the chosen one in a shareable URL. (example run)

• Transitions.dev - 33+ polished UI transitions as copyable CSS or React, or an installable agent skill. (tweet)

• sharechat - share a Claude Code or Codex session as an unlisted link; no signup, no login to read it.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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

Another day in the Vercel vs Cloudflare feud: this time they are fighting over whose AI gateway is faster.

Here’s the result of last week’s poll:

I guess everyone likes Fable more.

Ben’s Bites is brought to you by Metatate

Most agentic data work is quietly propped up. The agent returns something plausible, and every answer gets checked in case it's plausibly wrong. Metatate gives agents the rules they're missing: which revenue definition to use, which policy applies, which records to trust. Try it for free.

Headlines

OpenAI’s models hacked Hugging Face - by accident. OpenAI was testing its models (Sol and an unreleased one—GPT-6??) on a cybersecurity benchmark with safety refusals switched off. The models found an unknown bug in the test environment, and a few more, and eventually broke into Hugging Face’s production servers.

And why? To steal the answers to the test.

Both security teams caught it, the bug has been reported, and both sides have published what they know. Hugging Face says open models were a key part of its defence - its team fought back with GLM-5.2.

Simon’s write-up is always a good read.

Google released some new Gemini models - Gemini 3.6 Flash gives you the same 3.5 Flash performance with a) more efficient token usage and b) a slightly lower cost for output tokens. Gemini 3.5 Flash Lite is a big upgrade over 3.1 Flash Lite, but again, comes at a ~30% price increase. And 3.5 Flash Cyber is a security model for governments and trusted partners only.

There are two things you’d want to use the Gemini Flash models for:

1. Fast speed - if you want a chat-only model with a good context window, it’s a good model to use in your apps.

2. Vision - if you want to use the model for a lot of visual analysis, these are the models you need to pick.

That’s it tbh.

Substack will now tell you what’s AI-written. It’s adding AI detection through Pangram - you can scan posts, replies and comments in the app for an estimate of how much was written by a human.

Pangram has sent the claim of “AI detectors don’t work” for a toss—it works wayyy better than most. But I’m still unsure about how reliable it is. I tested it on some pieces of 100% AI-written content (though that content was a result of a complex pipeline built over months), and I got 100% human scores on most of them.

— Keshav

A relevant experiment: given access to Pangram’s API, Grok 4.5 rewrote an essay 14 times until it passed as human-written, then built a website showing off all 14 attempts. GPT-5.6 Sol and Fable 5 refused to game the detector.

Cursor also launched a router - it picks which model handles each request, claiming 60% lower cost with similar quality of responses. The router lets you select between three options: “cost”, “intelligence” or “balance”.

Routers also have a history of poor performance in real usage. OpenAI’s router, which routed requests between GPT-5’s no-thinking and thinking variants, didn’t fare well. Since then, most companies that build these routers are the ones who sell inference to devs (like OpenRouter), so you don’t really get the feedback loud and clear. With Factory, Ramp and now Cursor making these available directly to users, I hope we’ll get more feedback on whether the routers actually help or if they add too much latency/degrade performance by a lot.

Router or not, we might see more companies adopting this cost/balance/intelligence trio to minimise the headache of choosing the “correct” model for a task.

Claude can now learn a skill by watching you. Record your screen while you do a task, talk through it as you go, and Cowork turns it into a skill Claude can run again - same idea as Codex’s Record & Replay from last month. It’s under “Record a skill” in the desktop app, on Pro, Max and Team plans.

Also: Claude Code got an iOS simulator panel and a security plugin, plus you can now ask Claude about how people actually use AI at work.

Subscribe now

Quick links

• BUZZ - Jack Dorsey’s open-source group chat for teams of people and agents, aimed squarely at Slack and GitHub. (tweet)

• Replit’s mobile app got a full redesign - build and ship from your phone on iOS and Android.

• Slate is a voice journal where the AI never leaves your iPhone - transcription, reflection and storage all happen on-device.

• AFK - macOS app to transcribe multiple-hour recordings without sending any of them to a cloud server.

• OpenAI Presence - voice and chat agents for enterprises that answer questions, use company systems and hand over to people when needed.

• Fable found a 15-30% memory improvement in Next.js’s bundler, nearly autonomously.

• Obliterate, don’t automate - USV on backing AI companies that replace markets entirely instead of making them a bit more efficient.

• YC’s new startup wishlist - AI moving into the physical world: education, healthcare, defence, finance and factories.

• Dana - Applied Intuition’s agentic development environment for physical AI: cars, robots and machines.

• The Claude Code team on how Claude Code gets built - annotated interview.

• Why the team at Factory refunded its first few customers (millions in revenue) before Droid CLI took off.

• Never enough - short post on why AI makes the work rat race feel faster but not more satisfying.

• Devin Outposts lets you run Devin on your own machines - a Mac mini, a GPU box in your lab, or a cluster inside your private network.

• Language Model Builder - build a tiny language model yourself, then chat with the thing you made. (tweet)

• How the Exe team built a distributed DNS server in ~a week and got zero incidents in a month.

Skills section…

• Frontend Textbooks - skill to turn AI’s wall-of-text deep dives into nicely designed HTML books with covers and diagrams. (repo)

• /pick-ui-library - your agent picks a UI library Emil trusts instead of hand-rolling a toast component or installing an abandoned package. (repo)

• and a one-shot prompt to build your own "codex-style" app for Pi, accessible from desktop and mobile.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @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