Overview Treat each run as both execution and research. Capture reusable successes and failures locally, promote only repeated cross-project patterns globally, and require human review before a lesson becomes a universal rule.