Cognition applies GPT-6 Astra to Devin’s software testing

Cognition is deploying GPT-6 Astra across Devin, its autonomous software engineer, as well as its command-line and desktop products. The company says the model improves Devin’s ability to test software and demonstrate that its changes work, with the aim of making code review more efficient as Cognition’s engineering teams produce more code.

Walden Yan, Cognition co-founder, said a major improvement from Astra is its ability to test work and prove that it functions as expected. Cognition uses Devin with businesses ranging from large banks to technology-native startups, making the quality and traceability of automated changes central to the product’s use.

Testing output includes recordings and coverage reports

In one example, Devin used Astra to test Otter Run, an iPhone game. It returned a recording of the game running in a simulator alongside a report that identified which checks had passed and which areas remained untested.

The recording provides engineers with a view of the application’s behaviour, while the report documents the scope of testing. Together, those outputs are designed to let reviewers inspect how the software behaves and identify work that may still require attention rather than relying solely on a code change.

Bug reports can be returned with visual evidence

Cognition is also using Astra when customers submit screenshots of bugs. The team can pass a screenshot to Devin, which fixes the issue and returns a screenshot showing the resulting state. Yan said this workflow is helping Cognition respond to customers more quickly.

The deployment follows Cognition’s broader push to expand Devin’s capabilities while the company pursues growth, including Cognition's high-valuation funding talks that highlights investor interest in the autonomous software engineering business. Astra is being applied across the product line rather than being limited to a single testing demonstration.

Review remains focused on evidence and exceptions

Cognition expects that stronger testing and visible evidence could eventually reduce the amount of code engineers need to examine manually. Yan said the company expects to look at less code over time and ship more, but the examples described still preserve a role for engineers: reviewing recordings, test reports and explicitly untested areas.

For businesses adopting AI-assisted development, the practical implication is to treat automated fixes as reviewable deliverables: require reproducible test evidence, identify coverage gaps clearly and keep human attention on the changes that evidence does not resolve.