Self-Harness

It operates in a three-stage loop: (1) a weakness-mining stage that analyzes execution traces generated as the agent performs tasks to identify failure patterns, (2) a harness-proposal stage that suggests modifications to the harness code to address those weaknesses, and (3) a proposal-validation stage that runs regression tests to ensure the proposed changes do not degrade existing performance.