Built-in tools
Knowledge and reasoning
knowledge_retrieval, faq_retrieval, web_crawl, rewrite_and_think, and conversation_memory.Tickets
create_ticket, get_ticket_status, update_ticket, and close_ticket.Contacts and identity
seek_contact, update_contact_profile, and verify_email_otp.Workflow
escalate_to_human, mark_query_resolved, save_unanswered_question, and send_email.Execution path
The tool utility layer caps repeated calls to prevent runaway loops. Tool events can be streamed to the console for transparency and stored as agent-run steps.Add a tool safely
1
Define one narrow action
Give the tool an unambiguous name, a precise description, and the smallest useful argument schema.
2
Validate at both boundaries
Validate model arguments in the agent and independently validate/authenticate the internal gateway request.
3
Scope by tenant and conversation
Never let the model choose an unrestricted organization identifier. Derive context from the job and verify records belong to it.
4
Design idempotency and confirmation
Make repeatable writes safe. Require explicit user confirmation for destructive, financial, identity, or externally visible actions.
5
Record useful telemetry
Capture duration, success, a redacted result summary, and a stable error category without logging credentials or personal secrets.