ExecCritic

A Microsoft Research paper (with UW–Madison and Georgia Tech) on the failure mode where a coding agent writes both the patch and the test that checks it, so their errors agree and produce false confidence. ExecCritic is a test–verify–revise scaffold that separates the roles: a Test agent generates repository-native tests, a fail-closed harness qualifies and freezes them, and a Repair agent revises source code from their execution feedback without touching the tests. Each role is a separately trained Qwen3.5-35B-A3B policy with its own reinforcement-learning recipe: "Learn to Test" rewards tests that distinguish correct from incorrect patches, and "Test to Improve" trains the repair agent on direct resolution and feedback-guided revision. Composing the trained agents reaches 72.6% on SWE-bench Verified, 11.4 points above the paper's baseline. Filed for the harness line: it is a scaffold result, not a base model result.