Over five weeks, participants build a harness against a real brownfield codebase, with least-privilege permissions, characterization tests and sensors, independent review, and CI checks for unattended agents.