Bisect is a self-improving agent harness that never debugs the same bug twice. Agents that learn rules from their mistakes also learn bad habits. Bisect runs an autopilot loop around a customer-support agent: the agent writes its own rules, each is tested before going live, and live behavior is monitored. When something breaks, Bisect searches the agent's rule history, verifies the culprit with fresh runs with and without it, and removes only that rule, keeping everything else the agent learned. Each conviction becomes an "antibody" in MongoDB, so if the agent tries to re-learn the same habit in new words, Atlas Vector Search recognizes it and a quick replay blocks it. Vercel Link: https://mongodb-hackathon-sept26-bisect.vercel.app/#7
Team
Placement
Participant
