Identify the failing stage#
Record the workspace, agent ID, run ID when available, time, call direction, channel, and exact error. Remove secrets before sharing logs.
| Symptom | Check first | Reference |
|---|---|---|
| API returns 401 | API key, header, and owning organization | Authentication |
| API returns 402 | Feature entitlement, workflow limit, or inference balance named in the error | Billing |
| Draft will not publish | Validation errors, complete graph, and conflicting triggers | Versions |
| Call starts without audio | Microphone, model validation, WebRTC/ICE, and TURN | Setup troubleshooting |
| Inbound number reaches the wrong agent | Number mapping, provider webhook, and published trigger | Inbound routing |
| Outbound initiation succeeds but no conversation occurs | Carrier events, destination restrictions, caller ID, and media | Telephony |
| Tool is never called | Node attachment, prompt instruction, schema, and required values | Tools |
| Knowledge produces no answer | Document processing state, embedding configuration, and query relevance | Knowledge base |
| Webhook did not update your system | Completed run, template values, credentials, receiver logs, and retry policy | Webhook payloads |
| Production differs from an editor test | Published definition, runtime context, and model overrides | Testing |
Change one responsible part#
A healthy homepage does not establish that model calls, media, storage, or a queue worker work. Reproduce the smallest failing operation, inspect its evidence, fix that boundary, and run the same check again.
Escalate with useful evidence#
Provide the exact failing step, expected and actual result, timestamp, relevant IDs, and redacted error text. Include whether text chat, Web Call, telephony, and downstream delivery passed independently. For self-hosting, include the deployed image or commit and the affected service logs.