article --summary
The summary
The article looks at how Anthropic verifies code that its AI agents write and review, without requiring engineers to read every change. Anthropic describes having Claude run changes inside live applications, using repository-specific verification instructions so agents can check their own work before committing.
Cost and trust are the central concerns. Anthropic says it runs cheap deterministic checks such as builds, linters, and unit tests before invoking model review, and it sets rules so an agent cannot modify the test that judges its own change. Commit gates, bug-derived reliability checks, and logs of how each commit was verified are also covered.
The company says merged pull requests more than doubled between June and September while reverts fell by over 40%, though it did not isolate how much of that came from verification. It grades its reviewers against past incidents, but has not published their false-positive or false-negative rates, and the piece notes that engineers remain accountable for every shipped change.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.