AgentRunbook: production postmortems for AI agents

Writes real incident root causes back into coding agents' memory

Problem

An AI agent writes code, tests pass, it deploys, then production breaks. A human fixes it, but the agent never learns what went wrong and tends to repeat the same mistake in the next session. Existing tools compare deploys but do not turn 'production behaved differently than expected' into agent-readable feedback.

Solution

After deploy, tap into app health metrics and error logs. When anomalies appear, pull the related code changes and generate an incident card: expected behavior, actual behavior, root cause, fix. Serve those cards to agents via MCP or rule files so the next session reads past lessons first.

Users

Indie developers and small teams using AI coding agents; engineering leads wiring coding agents into production.

MVP

Link deploy versions to error/metric anomalies; auto-extract related changes and generate incident cards; store and search cards by repo/module; expose cards to agents via MCP at session start; human confirmation before archiving.

Revenue

Monthly subscription by repo count; free tier keeps the latest 20 incident cards; team plan per seat.

Why now

Coding agents in production are a 2026 reality, yet the loop where agents learn from production feedback is almost unbuilt; a solo developer can ship a thin, accurate version first.