Cognition's Devin: The AI Coder That Now Checks Its Own Work

Cognition, the company behind an AI coding assistant called Devin, says its tool can now check its own homework. Devin writes software code, and until now a human engineer had to sit down afterwards, read every line, and figure out whether it actually worked. With a newer AI model called GPT-6 Astra plugged in, Devin can run tests on its own code and show proof that it functions, not just claim that it does.

Why does that matter to you? Because the slow part of AI-written code was never the writing — it was the reviewing. Think of a contractor who not only builds you a shelf but hangs a heavy box on it to prove it holds. That's the difference here. Engineers currently spend a huge share of their week reading other people's code and hunting for mistakes. If an AI can hand over code plus evidence it works, that review time shrinks, and software gets finished faster. Faster software means cheaper products, quicker fixes when something breaks, and smaller teams doing more.

The bigger trend is what the industry calls "agents" — AI that doesn't just answer questions but takes actions, like writing files, running programs, and checking results. Coding agents are the leading edge of that shift, because code is easy to test: either the program runs or it doesn't. That makes software the first place where AI can genuinely be trusted to finish a job rather than start one.

The catch is real, though. An AI checking its own work is a bit like a student grading their own exam — it tends to miss the mistakes it didn't anticipate. Devin can confirm the code runs, but "runs" is not the same as "safe, secure, and does what the customer actually wanted." Human engineers will still review the important decisions and the risky parts. The announcement also stays vague on numbers: we don't know how much review time drops, how many bugs slip through, or what it costs. Treat the promise as promising, not proven.