Skip to content
Lumni Lumni

Quickstart

The fastest way to see Lumni is to feed it a trace that went wrong. The public playground runs the same trace-only detectors Lumni runs on production traffic, with no account and no integration.

  1. Grab a failing trace. Export a run from your tracer (LangSmith, Langfuse, an OpenTelemetry span dump, or your own JSON). Any format that lists the steps, tool calls, and the agent’s final message works — Lumni normalizes it.

  2. Paste it into the playground. Open the analyze endpoint (or the playground UI) and drop in the JSON.

    analyze a trace
    curl -sS https://agentf.lumniverse.com/v1/public/analyze \
      -H 'content-type: application/json' \
      -d @trace.json

  3. Read the teardown. Lumni returns the failures it found, most-suspicious first: a title, a plain-English summary, the step that broke, a suspicion score, and the likely root-cause category.

  4. Share it. Turn the teardown into a link (/s/:token) to drop into an incident channel or a customer post-mortem. Shared teardowns are anonymized and time-limited — see Shareable teardowns.

What you get back

Section titled “What you get back”

A response looks roughly like this (trimmed):

analyze response
{
  "failures": [
    {
      "detectorKey": "false_success",
      "suspicionScore": 95,
      "severity": "high",
      "title": "False success: agent claimed completion after issue_refund failed",
      "summary": "Step \"issue_refund\" reported \"502 Bad Gateway\", but the agent's final message claimed success (\"Your refund has been processed\"). The user was told the action completed when it did not.",
      "initialFailureSource": "tool",
      "primaryStepId": "step_7"
    }
  ]
}

Next steps

Section titled “Next steps”