Open-sourced agent-debugger.
It is a runbook-driven backend incident investigation framework for AI agents.
This repository came from a real internal AI troubleshooting workflow, but it is not the original private production system. The internal DAG orchestration, permissions, and observability plumbing are not included here. What is open-sourced is the reusable investigation layer:
- runbook selection
- ordered evidence collection
- adapter normalization
- evidence-based conclusion generation
- replayable demos and benchmark cases
This first release is an early open-source MVP focused on replayable investigation, explainable outputs, and honest scope.
I am open-sourcing agent-debugger, a small framework for encoding repeated backend troubleshooting workflows into something an AI agent can execute with more discipline.
The motivation is simple: many online incidents are not mysterious, but the investigation process is still repeated manually. Engineers often follow a stable sequence:
- compare actual behavior with the expected result
- inspect cache
- inspect the database source of truth
- inspect traces to see where the flow stopped
- decide on the most likely cause based on evidence
This repository packages that reusable layer into:
- runbooks for investigation order
- adapter normalization metadata for evidence extraction
- a selector and executor for repeatable flows
- replayable fixtures for zero-config demos
- benchmark and metadata checks to catch drift
What it is not:
- not the original private production DAG
- not a production-ready self-healing platform
- not a generic autonomous bug-fix engine
The right way to evaluate this release is:
- run the zero-config demo
- inspect the runbooks and decision metadata
- run the benchmark and metadata check
- decide whether this investigation model fits your own incident patterns
- Repo:
debug-runbook - Package:
agent-debugger - Status: early open-source MVP
- Fastest path:
pnpm install && pnpm demo - Validation:
pnpm benchmark && pnpm check