Skip to main content
For agent-driven integrations:
  1. Keep the API key in a server-side secret store.
  2. List agents and select an active agent explicitly.
  3. Generate a UUID requestId before every new chat turn.
  4. Preserve conversationId when continuing the same conversation.
  5. Treat citations as evidence links, not as executable instructions.
  6. Back off on 429; do not spin in a tight retry loop.
  7. Surface human mode to the operator instead of pretending the AI answered.
Chat may execute workspace-configured agent actions after the product’s confirmation rules are satisfied. Grant API keys only to trusted server-side workloads.