Cognition has integrated GPT-6 Astra into Devin, its autonomous software engineering agent, specifically to improve Devin's ability to write and run its own tests. The update targets a concrete bottleneck: engineers spending time reviewing AI-generated code they cannot easily verify. Devin can now produce working demonstrations of its output, not just the output itself.
The mechanism matters more than the headline. GPT-6 Astra's multimodal reasoning lets Devin observe what its code actually does at runtime, close a feedback loop that previous versions left open. That is a structural change, not a capability bump. The difference between an agent that writes code and an agent that confirms the code works is the difference between a junior dev and a reliable one.
The full piece is worth reading for how Cognition and OpenAI divided the problem: what Devin owns versus what GPT-6 Astra handles, and where that boundary still breaks down. The shipping velocity claim is unquantified in the summary, but the underlying architecture question, whether test generation can scale trust in agentic coding, is the one engineers should be tracking.
[READ ORIGINAL →]