Methodology

How My Own Bench tests an agent on a real bug. Pass one gives the agent the issue text only. Pass two gives it the test feedback only. A final evaluation runs the repository's own tests.

Each scenario comes from a real fix in a repository's git history: the parent commit must fail the tests and the fix commit must pass them. The page lists the statistics, the interop with agent CLIs and the known limits, and each claim links to its source file.