The problem
AI agents are hard to trust in production because teams can’t see plans, tool calls, failures, retries, costs or approval decisions. When an agent produces a wrong or risky answer, there’s rarely a way to inspect how it got there.
How it works
Every run is recorded as a trace of typed steps. A dashboard shows metric cards, a runs table, run details and a trace replay, so you can see which tool failed and retried, estimate cost, classify risk and hold risky actions for human approval.
The demo agent solves a realistic SRE task end to end: analyze a failed API incident and write a safe runbook. A typed state flows through nine nodes: planner, log analyzer, metrics analyzer, root cause, runbook, risk classifier, approval, critic and audit.
Traces are stored in SQLite by default, or PostgreSQL via docker-compose.
03 /
Key decisions
- A small graph runtime instead of LangGraph
Nodes are pure state-to-state functions over a typed state object. The demo stays deterministic and easy to explain, and swapping in LangGraph later is straightforward.
- Every step is a typed record
Each node logs an AgentStep, so a run can be replayed, costed and audited after the fact.