Skip to content
Customer Pro
Start free

AI agents

Custom tools

REST and database calls an agent can make mid-conversation to fetch or change real data.

On this page

Without tools, an agent can only talk about what is in its prompt. With them, it can look up a real order, check real stock, and write a real record — mid-sentence, while the customer waits.

Audio
Letting an agent do things, not just talk

When you'd use it #

Order status

"Where is my order?" becomes a real answer from your own system instead of an apology.

Availability

Check stock or a booking slot before the agent promises anything.

Capturing a lead

Write the enquiry into your CRM as it happens, rather than reading it back out of the transcript later.

Defining a tool #

  1. Name it for the model, not for you

    get_order_status tells the model when to reach for it. tool_2 does not. The name and description are how the model decides.

  2. Describe when to use it

    One clear sentence. This is the single biggest factor in whether the tool gets called at the right moment.

  3. Declare the parameters

    What the model must supply, and their types. Keep the list short — every parameter is something the model can get wrong.

  4. Configure the call

    For a REST tool: method, URL, headers and how parameters map into the request. For a database tool: the query and its bindings.

  5. Attach it to an agent

    A tool exists here but is only usable by agents you attach it to.

Keeping it safe #

Field What it does
Prefer read-only Most conversational questions need a lookup, not a write. Reserve writes for tools you have watched behave.
Scope the credentials Use a token that can reach only the endpoint the tool needs.
Never return more than the answer Whatever the endpoint returns can end up paraphrased to a customer. Do not hand back the whole record.
Test in the playground Watch which arguments the model actually invents before it does so in front of someone.

Where to go next #

Agent builder Persona, model, temperature, tools and the WhatsApp numbers an agent answers on. Playground Test an agent against real prompts before a customer ever reaches it. MCP servers Connect Model Context Protocol servers to hand an agent a whole external toolset at once. Audit logs Every inference an agent made, with status, tokens, latency and the tools it called.