AutoFix AI
A 6-agent autonomous DevOps incident-response command center that investigates production error traces, evaluates fixes in a self-correcting feedback loop, and drafts pull requests.
Project overview
AutoFix AI is an autonomous DevOps incident-response system built with six specialised AI agents orchestrated over a Node.js + Express backend. When an error or incident is reported, the agents collaborate to investigate, select and validate a fix, then produce a complete incident report and communication draft.
The problem
Debugging a production incident requires searching documentation and forums, evaluating multiple candidate fixes, validating the chosen fix, writing an incident report and drafting stakeholder communication—each step consuming engineering time that could be redirected to prevention.
Six-agent pipeline
Manager Agent
Creates a structured investigation plan when an incident is reported.
Researcher Agent
Searches documentation, Stack Overflow and GitHub issues for relevant solutions.
Analyst Agent
Evaluates the found solutions and selects the most applicable one.
Tester Agent
Validates the proposed solution against the error context. Returns pass or fail with reasoning.
Writer Agent
Produces a structured incident report including root cause, solution and action steps.
Action Agent
Drafts external communication (status update) and pull-request description content.
Self-correction loop
When the Tester Agent rejects a solution, its failure reasoning is fed back to the Researcher Agent, which searches for an alternative. The cycle continues until a valid solution is found or the attempt limit is reached.
Technical architecture
Frontend
- React (Vite)
- Real-time agent status updates
- Incident report viewer
Backend
- Node.js + Express proxy
- Multi-provider AI (Anthropic, Gemini, OpenAI, Groq)
- Agent orchestration & self-correction loop
- Structured output parsing & PR diff generator
Challenges & trade-offs
Termination condition. Without a bounded retry limit, a self-correcting agent loop can run indefinitely. The system limits correction iterations and falls back to a best-effort report if no valid fix is found within the budget.
Multi-provider proxy & rate limits. AutoFix AI proxies across multiple LLM providers (Anthropic, Gemini, OpenAI, Groq). When a primary model experiences latency spikes or quota exhaustion during an incident, the dispatcher routes the investigation plan to secondary fallbacks to preserve uptime.
Honest future improvements
- Add tool-use so the Researcher Agent can execute safe read-only terminal commands
- Connect to real monitoring alerts (e.g., PagerDuty, Prometheus) as input sources
- Store incident history for pattern analysis across repeated failures
- Add configurable retry budget and escalation thresholds