Cognition has Devin use GPT-6 Astra to test and prove its own code changes

Cognition, maker of the autonomous coding agent Devin, is now using OpenAI's GPT-6 Astra to have Devin verify its own work before handing changes back to human engineers — closing a loop that previously required a person to manually check whether an AI-written change actually worked.

What's new

Walden Yan, Cognition's cofounder, said the shift is specifically about proof, not just code generation: "One of the big pieces that Astra improves on is its ability to test and prove that its work actually functions the way you expect." The integration spans Cognition's product line — the core cloud agent, the CLI, and desktop products.

In one demonstrated case, Devin used Astra to test an iPhone game called Otter Run, returning both video recordings of the test runs and detailed reports showing which checks passed and which parts of the app remained untested. In another workflow, when customers report bugs via screenshots, Cognition deploys Devin with Astra to fix the underlying issue and return screenshots proving the fix, rather than just a diff.

Context

Devin has been marketed since its debut as a step toward autonomous software engineering, but the persistent bottleneck for agentic coding tools has been verification: an agent that writes plausible-looking code still needs a human to confirm it actually does what was asked. Cognition's pitch here is that Astra's stronger tool use and evaluation ability lets Devin close that verification gap itself.

Why it matters

Yan framed the payoff in blunt terms: "We expect over time that we have to manually look at less code and end up shipping more at the end of the day." That is the central promise of agentic coding tools — not just writing code faster, but reducing the amount of human review required per unit of shipped work — and Cognition is one of the more visible vendors betting its product roadmap on a specific underlying model's ability to deliver it.