Incident Commander

AI-powered incident response agent — built for OpenAI Build Week 2026.

When a production incident hits, most engineers spend the first 20-30 minutes just figuring out what's going on — digging through logs, checking dashboards, tracing which service actually broke. Incident Commander tries to shrink that window. You describe (or simulate) an outage, and a pipeline of specialized AI agents investigates it the way a senior SRE would: checking health metrics, searching logs, tracing service dependencies, and pulling relevant runbooks — before handing you a diagnosis and a fix.

How it works

Four agents run in sequence, each one handing its findings to the next:

Alert → Triage Agent (health_check, metric_query)
            ↓
       Diagnosis Agent (log_search, service_map, metric_query)
            ↓
       Fix Agent (runbook_search, service_map)
            ↓
       Postmortem Agent (full context)
            ↓
       Complete incident writeup

Each agent isn't just prompting an LLM once — it runs a real tool-calling loop. The agent decides what to check, calls a tool, reads the result, and keeps investigating until it has enough evidence to reach a conclusion with a confidence score attached.

Features

• Multi-agent pipeline — triage, diagnosis, fix, and postmortem are handled by separate agents with their own system prompts and tool access, not one giant prompt trying to do everything
• Real tool use — 5 function-calling tools (log search, metric query, service map, health check, runbook search) that agents call to actually gather evidence instead of guessing
• Structured outputs — every agent returns a strict, typed schema (severity, evidence chain, confidence score, etc.), not free-form text
• Live streaming — results stream to the frontend in real time over WebSocket as each agent finishes its stage
• One-click demo — a pre-built outage scenario you can trigger instantly for a clean demo run

Tech stack

Component	Technology
Frontend	Vanilla HTML/CSS/JS (single-page, no build step)
Backend	Python, FastAPI, WebSocket
AI	OpenAI GPT-5.6 (function calling + structured outputs)
Database	SQLite

Getting started

Prerequisites

• Python 3.10+
• An OpenAI API key with access to GPT-5.6

Backend

cd backend
python -m venv venv
source venv/bin/activate        # Mac/Linux
# venv\Scripts\activate         # Windows

pip install -r requirements.txt
cp ../.env.example ../.env      # then add your real API key to .env
uvicorn main:app --reload --port 8000

Frontend

No build step needed — it's a single static page. Either:

cd frontend
python -m http.server 3000

then open http://localhost:3000, or just open frontend/index.html directly in your browser.

The frontend also has a built-in demo mode that works without the backend running — useful for a quick, guaranteed-not-to-break demo.

Running the demo

1. Open the frontend
2. Click "Simulate Outage"
3. Watch the agents investigate in real time — triage, then diagnosis with evidence, then a fix with commands, then a postmortem

Project structure

incident-commander/
├── backend/
│   ├── agents/          # triage, diagnosis, fix, postmortem agents + orchestrator
│   ├── tools/            # the 5 function-calling tools
│   ├── models/           # Pydantic schemas for structured outputs
│   ├── services/         # vision, voice, incident services
│   ├── routes/           # API endpoints + WebSocket
│   ├── mock_data/        # simulated logs, metrics, and runbooks for the demo
│   └── main.py
├── frontend/
│   ├── index.html
│   └── styles.css
└── README.md

Built for OpenAI Build Week 2026

Category: AI Agents

Built with Codex on GPT-5.6 — multi-agent orchestration with real tool calling, structured outputs, and real-time streaming, wrapped around a scenario every engineer has lived through: the 3 AM outage nobody understands yet.

License

MIT