The builder is where an agent's behaviour is decided. Most of the screen is optional; the parts that are not are the prompt, the model, and the numbers it answers on.
Identity and prompt #
| Field | What it does |
|---|---|
| Name required | Internal. Appears in agent pickers, call routing and audit logs. |
| Description | A note for your team about what this agent is for. |
| System prompt required | The instructions the model receives on every turn. Say who it is, what it knows, what it must not do, and what to do when it does not know. Explicit refusals matter more than encouragement. |
| Variables | Named values you can reference from the prompt, so the same agent can be reused with different specifics. |
| Active | Off means it never answers. This is the safest kill switch. |
Model settings #
| Field | What it does |
|---|---|
| Provider and model required | Which connected provider and which of its models to call. |
| Temperature | How much the output varies. Low for factual support, higher for conversational sales. Start low. |
| Max tokens | A ceiling on reply length. Also a cost ceiling. |
| Top P, frequency and presence penalty | Finer sampling controls. Leave them alone unless you have a specific reason. |
| Memory window | How many previous turns are included. Larger means better continuity and higher cost per message. |
Capabilities you can switch on #
| Field | What it does |
|---|---|
| Knowledge base | Search your documents before answering. See Knowledge bases. |
| Tools | REST and database calls. See Custom tools. |
| Custom response | Force a typed output shape. See Custom responses. |
| Voice reply | Speak replies aloud, using a text-to-speech provider, and transcribe inbound voice notes. |
| Vision | Read images the customer sends — a photo of a damaged item, a screenshot of an error. |
| Contact allowlist | Restrict the agent to a named set of contacts. Useful while piloting. |
| Payments | Let the agent request money mid-conversation, with tiers and an amount ceiling you set. See Payments. |
Binding it to WhatsApp #
An agent does nothing until it is bound to at least one connected WhatsApp number. Binding is what puts it on duty for messages arriving there. One number can have one agent on reply duty; one agent can cover several numbers.
Where to go next #
Playground
Test an agent against real prompts before a customer ever reaches it.
Custom tools
REST and database calls an agent can make mid-conversation to fetch or change real data.
Knowledge bases
Documents an agent searches before it answers, so replies come from your material.
Template assignment
Decide which AI agents are allowed to send which templates.