FAQ
How is Lumni different from observability (LangSmith, Langfuse, Helicone)?
Section titled “How is Lumni different from observability (LangSmith, Langfuse, Helicone)?”Observability shows you the trace — what happened. Lumni tells you why it failed, whether a fix works, and blocks the same regression from shipping again. It runs detectors for silent failures, does root-cause analysis, replays fixes, and gates releases. It’s additive — keep your observability tool; Lumni reads the trace, it doesn’t replace the tracer. See the comparison.
Can Lumni interfere with my agents or move money?
Section titled “Can Lumni interfere with my agents or move money?”No. Lumni is read-only and advisory by design. It cannot issue refunds, send messages, change records, or modify agents. Gates fail open, so Lumni can never break a deploy.
Which frameworks does it support?
Section titled “Which frameworks does it support?”Anything you can express as a trace. Native paths include LangChain, LangGraph, CrewAI, the OpenAI Agents SDK, Retell, Vapi, and Zendesk AI, plus any OpenTelemetry / OpenInference source. If it produces a trace, Lumni can analyze it.
How much data do I need to send?
Section titled “How much data do I need to send?”Enough to be detective-grade, not exhaustive. The trace-only detectors work from step statuses, tool names/inputs, the final message, and a bit of metadata (model, tokens, cost). You can start by pasting a single trace — no ingestion at all.
What if my agent lives inside Zendesk or Intercom?
Section titled “What if my agent lives inside Zendesk or Intercom?”Use a read-only support-tool connector. Lumni reconstructs a run from the ticket transcript and recorded actions and runs the same detectors — which is how it catches a bot that says “refund processed” in a ticket with no matching refund action.
What’s a “suspicion score”?
Section titled “What’s a “suspicion score”?”A 0–100 triage signal, tuned per detector, for how confident Lumni is that a real silent failure occurred. It’s not a probability and not a guarantee — sort by it and set your own threshold. See core concepts.
Do I get charged per event?
Section titled “Do I get charged per event?”No. Ingestion and detection are free. You’re billed for investigations — the root-cause, replay, and gating work Lumni does that you couldn’t do yourself.
Is there a way to try it without integrating?
Section titled “Is there a way to try it without integrating?”Yes — the public playground / analyze API. No account, no SDK, rate-limited per IP, no PII captured. Paste a trace, get a teardown, share it.
What does “silent failure” actually mean?
Section titled “What does “silent failure” actually mean?”A failure where no exception was thrown, so evals and generic monitoring miss it: the agent claimed success but a step failed, looped without progress, fabricated data, skipped a required action, or overflowed its context window. Those five are exactly what Lumni’s detectors catch.