Multi-agent triage for a simulated bottling line
Four machines, six hours of 10-second sensor data, at most one injected fault. Three AI agents detect the problem, look up the likely root cause in maintenance manuals and write the shift handover report. A human approves the work order. All data is synthetic.
- Pick a shift
- Explore the data (computed live, no AI involved)
- Run the agents and watch them work
1 · Choose a shift
Reveal what was injected (spoiler)
2 · Line data live computation · no LLM
3 · Agent investigation
How it works
- Deterministic first. Statistics, alarm limits and an Isolation Forest find the anomalies. The agents call these as tools, so numbers are reproducible and the model only has to reason and write.
- Cost-aware by design. Agents pass short summaries to each other, tools return compact JSON, every run has token, step and time limits, and recorded runs replay for free.
- Human in the loop. Work orders are only proposals until a person approves them. Nothing here talks to a real maintenance system.
- Grounded. Root-cause claims must cite the maintenance manuals (for example MNT-101 §3). Those manuals are fictional example documents.