Tools under the hood

Shared execution#

In the workspace-tools preview, eligible agents use the same backend tool runner as the manual Tools interface. Access is assigned by the agent’s job. A content agent can use relevant writing or campaign tools; an analysis agent can use relevant inspection tools.

Tool access does not give the model permission to choose another tenant or arbitrary website. The calling workspace supplies that scope, and the backend enforces membership and execution permissions.

Inspect the result#

Agent runs are marked as agent-originated in the saved history. A successful tool run has a result you can reopen and inspect. Where a generated output cites that receipt, verify the underlying evidence and execution mode before accepting its conclusion.

A tool result is source material, not an instruction to ignore the current brief or approval rules. File downloads and generated plans do not count as completed external actions.

Limits remain#

Tool execution is bounded and rate-limited. AI-backed operations still need an eligible account and available service. A tool failure should be visible; it must not silently turn into an invented live result. Agents still need normal approval for consequential publishing or connected changes.