Tool execution

Tool Execution Approval adds human approval before an agent runs selected skills. A tool is a skill available to the agent.

To configure approval, open Manage Agent → Guardrails. Turn on Tool Execution Approval, then select the gear in that row.

Choose tools that need approval

The Edit Tool Permission screen lists every skill available to the agent. Each skill has its own approval toggle. A counter shows the number selected, such as “18 of 19 skills selected.”

The Edit Tool Permission screen.

• Enable for all skills requires approval for every current and future skill. It overrides individual toggles.

• Per-skill toggles require approval for selected skills only.

• Search skills filters a large skill list by name.

The list includes built-in and added skills. Built-in examples include AI Reasoning, Web Search, Knowledge Search, Generate Custom Skill, and Extract Document Data (Legacy). The available set varies by agent.

Custom skills appear automatically. With Enable for all skills on, newly added skills inherit the approval requirement.

Select Update Tool Permission to save changes. Select Cancel to discard them.

Saving changes with Update Tool Permission.

What users see

When the agent reaches an approved tool, it pauses before the call. The chat shows guardrail activity, such as “Running tool-permission guardrail for ‘Web Search’.” It then asks the user whether the tool should proceed.

The runtime approval prompt for a paused tool call.

• Allow runs the tool and continues the agent run.

• Deny blocks the tool call. The agent continues without it.

Choose an approval strategy

Require approval for tools that take external actions or access sensitive data. Individual approvals work well when only a few skills need review.

Use Enable for all skills when every tool call requires confirmation. This setting also protects skills added after initial configuration.

For the parent-level overview, see Guardrails.