Early impressions of Cognition's Devin

Devin's AI software developer is remarkable

Every so often I find a technology that doesn’t just work well, but feels like a glimpse of the future that’s sneaking up on the rest of the world. Two days ago, Devin gave me that feeling. Devin is a “collaborative AI teammate”. You can add him to your repo, invite him on Slack, and immediately start asking him to complete work.

I’ll walk through a workflow where he attempts to fix a bug we’ve had in our product for a while:

In response, Devin shares a plan for implementation:

First, Devin got stuck because he didn’t know how to authenticate. I told him to create his own workspace through our sign up flow and that guidance worked.

Two hours later I got this:

Holy crap! The PR went up, someone on the team approved it, and the fix was merged. This was admittedly a simple bug to fix, but Devin still did it perfectly. And without Devin, we would not have prioritized the bug anytime soon.

But here’s where it gets interesting. You know that clarifying question Devin asked above about how to authenticate? I told Devin to create his own account, and Devin automatically saved this as under his “Knowledge repository”. Devin learns on his own — I don’t anticipate having to provide that same guidance again.

This is a glimpse of how engineering will evolve. Right now, Devin needs my permission. He waits for instructions, then goes to work. But as these systems improve, we’ll trust them more. Eventually, they won’t just wait around for requests; they’ll be proactive. A new bug ticket shows up in Linear? Devin will start implementing the fix before I say anything. A new way to improve performance that Devin discovered in another codebase? Devin will just implement that too. Codebases will be self-optimizing.

Already, Devin can handle different kinds of input. I showed him a screenshot on Slack and asked for a copy change in the UI. It found the relevant code and fixed it. It’s not hard to imagine the next steps: upload a video of a bug (that was sent by a customer) instead of describing it in text, and Devin fixes it. Or maybe Devin records a short demo video of each change so we can see it in action without manually testing. These agents will be multi-modal on input and output.

Just this week I discovered Devin, made videos with Sora, and upgraded to ChatGPT Pro. Each experience was profound. I’m so excited about what’s next.