Skip to content

Tools

A tool is an operation an agent can request through its harness. A tool can read a calendar, search records, or submit data to a service. A tool’s input and output contract gives the harness a way to validate the request and return evidence to the run.

Expose a tool with a name, input contract, and an operation the harness can authorize. The model selects a tool and supplies arguments. Validate the arguments and resolve the current workspace binding before acting. Tool results return to the harness as data that a later step can use.

A Studio connector tool resolves through this requirement in the Meeting desk fixture. The operation ID must also appear on the step that invokes it.

{
"id": "calendar",
"label": "Google Calendar",
"context": "personal",
"providers": [
"googlecalendar"
],
"operations": [
"upcoming"
],
"optional": false
}
  1. Resolve the declared tool in the running context and check the caller’s grant.
  2. Validate or map its arguments, then classify the effect from the supported operation.
  3. For a protected action, hold execution for the configured review and verify the approved proposal at dispatch.
  4. Return the projected result or a clear blocked or failed status. Record a receipt when an external effect occurred.

A playbook may expose an operation through MCP. MCP discovery and input schemas describe the tool. The current workspace, binding and approval still govern a real call.

See connector operations, approvals, and MCP interoperability.