Agent infra

AgentOps

A control tower for tracing, debugging and auditing agent runs.

How it works
  1. Plan
  2. Analyze
  3. Root cause
  4. Runbook
  5. Risk
  6. Approval
  7. Audit

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.