spin-off
A safety-first self-improvement loop for AI agents: execute → track → analyze → auto-apply → auto-rollback on regression.