Knowledge and tools

Attach knowledge and tools to an agent to give it context and let it perform supported actions during a call. Make instructions clear about when an action should be used, and test the action with realistic inputs.

A knowledge source can contain text or be fetched from a URL. URL sources can crawl linked pages within configured bounds. Sources move through pending, ready, and failed states; check that ingestion has completed before testing an agent.

Sources can be available to all agents in a workspace or only to selected agents. Prompt-mode content is bounded to 24,000 characters per source. The model also supports a retrieval usage mode; confirm it is configured and supported by your deployment before depending on it.

FAQs and scenarios can be attached to agents separately. Keep source material current and review generated answers before relying on them in customer conversations.

The workspace supports webhook, MCP, and client tool definitions. A tool has a name, description, input schema, and a timeout. Webhook and MCP tools require an endpoint; credentials configured for a tool are stored encrypted. The default tool timeout is eight seconds and the maximum is 30 seconds.

Built-in call tools include hold, resume, end, transfer, consult, and scheduling actions. Availability depends on the provider and call runtime. A tool definition alone does not guarantee that a particular provider can execute it; verify the full call path before publishing an agent.