Pattern: keep the test writer and the fixer separate, whether they are people or agents

arXiv·medium signalXBlueskyLinkedInCopy link

The Datasette audit split test-writing from fixing between two people. ExecCritic (arXiv 2609.09133, September 8) measures why that matters for agents. When the base Test agent wrote the tests a Repair agent worked against, SWE-bench Verified resolution fell from a 61.2% no-test baseline to 57.3%, because a patch and a test from the same source share the same mistakes. Its scaffold freezes independently written tests behind a fail-closed harness before repair starts. In practice, don't let the agent that writes the fix also write the tests that approve it.