Enable Tool Permission HITL
Tool permission HITL is a governance workflow: decide which tools may run without review, which must ask a human, and which should never be exposed. Use the Admin Console first so the policy is visible next to the agent draft.
What you click
Section titled “What you click”
- Open Tools and identify the sensitive tool ids.
- Open Agents and choose the agent.
- On Tools, expose only the tools this agent should know about.
- On Plugins, enable the permission plugin.
- Add rules:
- Allow for safe read-only tools.
- Ask for tools that can spend money, write files, call external systems, or reveal sensitive data.
- Deny for tools this agent must never call.
- Click Validate.
- Preview one prompt that should run without asking and one that should suspend for review.
- Save when both paths are correct.
Rule checklist
Section titled “Rule checklist”| Question | UI check |
|---|---|
| Does the agent need to see this tool at all? | Tools tab allow/exclude selection |
| Should the tool be available but reviewed? | Plugins → Permission → Ask |
| Should the tool disappear from the effective catalog? | Plugins → Permission → Deny or exclude it in Tools |
| Are argument-sensitive rules needed? | Add a rule that matches the tool arguments, then preview that exact case |
| Did the policy change after save? | History and Audit Log |
Verify and operate
Section titled “Verify and operate”
After saving, run a real task or eval fixture that tries the sensitive tool. A correct Ask rule should suspend for human review in your client/HITL surface. If the model can still call a tool that should be blocked, tighten the Tools allowlist and the permission rule together.