What's new with Replit Agent v2
An overview of what’s new and better about Replit Agent v2.
Replit Agent v2 launched on February 25th and it’s pushed itself to the front of the AI coding tool pack in terms of capability.
Right now, it’s in early access mode, available to those with paid plans. To use it, you need to enable “Explorer Mode” in the ‘Preferences’ settings panel.
UPDATE: Replit Agent v2 is now available for all users by default.
Let’s explore!
A more autonomous and iterative agent
The new version follows the release of Claude Sonnet’s 3.7 model, which is much better at coding. But, according to Replit, Agent v2 isn’t just a hot swap of models—rather, it’s a ground-up rebuild, making it more creative and proactive.
When announcing the launch, Replit highlighted that Agent is more autonomous with its work, saying, “[at] each step, it forms a hypothesis, searches for the right files, and only starts making changes when it has enough information to get the job done.”
As you build, you’ll notice this in action. Agent v2 shows you checklists as it goes, telling you what it has done and what it’ll work on next. This gives you better insight into its thinking process and the chance to jump in and correct things if it’s going off track.
This has two major benefits. First, the model tends to implement changes and new features in a more complete way—no more “now implement it on this page too” follow-up prompts needed.
This, in turn, means Agent v2 gets more done per checkpoint, making it more cost effective.
No more error loops (hopefully!)
Sonnet 3.7 is a reasoning model, and this ability to reason has made Agent v2 less prone to getting into error loops.
Instead of a hit-and-hope approach to fixing bugs, Agent v2 knows to step back, study the problem as a whole and only apply a fix when it’s confident about the underlying cause. It’s also better at testing its fixes on its code. So you can be more confident that it's actually fixed things.
This update also comes with improved behaviour around checking logs and spotting errors early every time it relaunches. Overall, it’s now even easier for non-technical folks to use.
Better at design
Replit has put a lot of work into Agent’s design abilities, which was a definite weakness in v1.
Generally, Sonnet 3.7 is much better at treating well-designed web apps through its code. Gone are dated interfaces and messy layouts.
But what’s really cool is that Agent v2 now builds in real time. After approving a build plan, while Agent codes in the background, it will start putting together a design piece by piece for you on screen.
We’re still waiting on the ability to select specific elements on the screen and edit them—a feature that some other AI coding tools have. Maybe that’s to come in v3!
What are people saying and building?
The launch was well hyped on X by members of the Replit team, and a quick search on X will show you what people have been building and saying (a lot!).
Of course, there have been some teething problems—for this author included. Every so often Agent v2 will spend a few moments confidently telling you all the things it’s done, only for you to find out it either a) didn’t actually do them or b) they don’t work well.
But it’s early days, and considering the pace of development since Agent first launched in October 2024, we can expect things to improve quickly.
This post was written by Andrew.