Scar Tissue

Scar Tissue: Every agent failure becomes a tested upgrade to its own harness. An LLM agent takes a small grocery's orders in MongoDB Atlas. When a run breaks an invariant (the order API times out after saving, and the retry orders twice), the harness, with no human: Turns the incident into test cases; Recalls past incidents and their fixes with one $lookup; Asks a model for three typed policy changes: rules, guardrails, context or tool access; Drops any that name the incident, and tests the rest against a fixed evaluator it can't edit. Only a candidate that passes every case ships; Writes the winner to Atlas, and the agent reloads it through a change stream, pinned by hash. Agents that take real actions can't learn by trial and error in production. Scar Tissue makes the postmortem executable and tested. https://scar-tissue-five.vercel.app/