Incident Commander
A team of AI agents diagnoses a live incident in a simulated microservice stack and proposes a fix for you to approve.
Open the evaluation boardProblem
On-call engineers lose the first half hour of an incident collecting context from logs, metrics and deploy history.
Try it live
Break the shop. Ship a bad deploy, slow the payment provider, leak memory in inventory or start a cache stampede. Watch the dashboards turn red while the agents investigate, approve the rollback, and read the postmortem.
Proves
Multi-agent orchestration with budgets, real-time systems, and SRE practice.
Signal chain
- burn-rate alert
- commander plans
- logs, metrics + deploy agents
- rank hypotheses
- you approve
- verify SLO
- postmortem
Stack
Node · TypeScript · WebSockets · Redis Streams · event-sourced simulator · Vue dashboards
Engineering highlights
- Detection and correlation run as deterministic code and the agents reason over compact summaries, so an incident takes 9 to 15 model calls under a hard step cap.
- The simulator is seeded and event-sourced, so any incident replays exactly, for debugging and for replay mode.
- No remediation runs without a human click, and the incident closes only after the SLO recovers.
| Operating limit | Value |
|---|---|
| Incidents per visitor per day | 1 |
| Step cap per incident | 15 |
| Remediation | Human approval |
| Model calls per incident (est.) | 9–15 |
Status The evaluation board for this part is open. It starts on curated samples, and your own text runs live within your daily allowance. Open the evaluation board