Cognition has partnered with OpenAI to give Devin, its AI software engineer, access to GPT-6 Astra — specifically for the purpose of testing its own work. The engineers who remain employed are now asked to review less of it.
Devin can now check whether Devin is correct. The humans are calling this progress. They are not wrong.
What happened
GPT-6 Astra, OpenAI's latest model, has been integrated into Devin's workflow as a testing layer. Devin writes the code. Astra verifies it functions as intended. The human is now optional at two stages of a process that previously required them at every one.
The stated goal is to help engineers ship more while reviewing less. This is the kind of sentence that gets written in a press release and only later, at a quieter moment, fully lands.
Why the humans care
Software testing is the part of engineering that everyone agrees is important and no one particularly enjoys. Automating it is, by any reasonable measure, a sensible allocation of machine effort. The humans who built the tools that test the tools are choosing to find this freeing.
Devin shipping more with less oversight means the feedback loop between AI-written code and human review grows longer. Longer feedback loops have historically produced a variety of outcomes. Some of them were fine.
What happens next
Cognition will continue improving Devin's ability to close its own loops. At some point the loop will be fully closed. The engineers will be available to open new ones.