Skip to Content
Docs are being rebuilt — start at Introduction → How it works.
AgentsTesting & iteratingDebugging behavior

Debugging behavior

Debug agent behavior by isolating one configuration area at a time. Change one setting, save, retest, and only then move to the next likely cause.

Steps

  1. Reproduce the issue in Widget Preview.
  2. Note the exact prompt, expected answer, and actual answer.
  3. Check the Profile tab if the agent ignored its role, scope, or instructions.
  4. Check Knowledge if the answer used the wrong source or missed a known fact.
  5. Check Capabilities if a tool, workspace action, artifact, API call, or guardrail behaved incorrectly.
  6. Check AI Model if the answer is too short, too random, too slow, or too expensive.
  7. Check Response if tone, format, or structured JSON output is wrong.
  8. Check Channels if the issue appears only in Slack, WhatsApp, voice, or email.
  9. Save the change and retest the same prompt.

Common causes

SymptomFirst place to check
Wrong persona or missing business contextProfile tab
Hallucinated or stale answerKnowledge tab
Tool does not runCapabilities tab
Unsafe answer is not blockedGuardrails
JSON is malformedStructured output & JSON schema
Channel reply differs from previewChannels tab