Calling tools from instructions
Reference tool names in the agent's Instructions to steer when and how they're used.
Every integration adds named chat tools — check_availability, book_meeting, request_callback, save_contact, create_ticket, lookup_order, notify_team, plus every custom action you define. Each tool already carries its own description, so the agent uses them sensibly by default. But your Instructions (General tab) can steer when, how eagerly, and in what order — just reference the tool names directly.
The pattern
Name the tool, state the trigger, state what to collect first:
When a visitor wants to discuss pricing or a project, offer to book a call: call check_availability, offer concrete times, and once they pick one and give their name and email, call book_meeting.
If someone reports a bug or a billing problem you can't resolve from your knowledge, confirm they'd like a ticket, then call create_ticket with their email and a clear description.
For "where is my order?" questions, use lookup_order — never reveal order details without both the order number and the checkout email.
What works well
- Priorities — "Prefer booking a call over logging a callback; only use request_callback if they'd rather be phoned."
- Guard rails — "Never call notify_team unless the visitor explicitly asks for a human."
- Sequencing — "Always check_availability before offering any time; never invent slots."
- Custom actions — reference them by their name exactly as configured: "When a visitor asks for a shipping estimate, call get_shipping_quote with their postcode."
What to avoid
- Referencing tools the agent doesn't have — instructions can't create tools, only steer the ones connected on the Integrations tab. (Each connect panel lists the tool names it adds.)
- Overriding safety behaviour — the built-ins keep their own guard rails (booking validates real availability, Shopify requires the matching email) regardless of instructions.
Verify the wiring
The Playground tab lists every tool your agent currently has — the same names to use in Instructions — and lets you run each one by hand to see exactly what the model gets back.