Operations

Troubleshoot a deployment

Start with the boundary closest to the failure:

  1. Confirm the server loaded the intended CONFIG_DIR and agent key.
  2. Check provider credentials and GET /api/v1/models.
  3. Reproduce through the HTTP API before debugging a channel or UI.
  4. Inspect run events and tool audit entries.
  5. Test external tools, MCP servers, Redis, storage, or voice providers independently.

Common symptoms:

  • Agent missing: registration path, key, or config directory is wrong.
  • Tool missing: the tool set is not registered by the server or not granted to the agent.
  • RAG empty: the document set was not ingested or the source name differs.
  • Requests unauthorized: issuer, audience, key, or downstream scope does not match.
  • Files disappear: output remained in a sandbox instead of being published as an artifact.

Use for exact configuration and when the failure crosses subsystem boundaries.

Next

  • FAQ - the same problems, indexed by the message you are seeing.
  • Observe runs - traces and audit entries for a specific run.
  • - when a failure crosses subsystem boundaries.

Evaluate agent quality

Offline datasets before you ship, sampled scoring of real runs after.

Use cases

Three projects built end to end, from an empty config directory to something you can hand over.

On this page

Next