ImprovementJuly 3, 20261 min read
Selective tool approvals
Human-in-the-loop approvals get more precise and more dependable: gate only the tools you choose, enforce the same policy on every execution path, and let approvals wait durably for a decision.
agentssecuritytoolsdashboard
Human-in-the-loop tool approvals are now more precise and enforced everywhere your agent runs. You decide exactly which tools pause for a person, and that policy holds no matter how the agent is invoked.
- Approve only what matters. The agent editor's Safety section now offers a scope switch: require approval for all tools, or only a selected list. Pick from the agent's attached tools or add custom wildcard patterns like
mcp:*. - One policy, every path. An agent's approval config is enforced consistently across direct execution, dispatch, and client-token chat embeds, so an approval gate behaves the same on your website widget as in the dashboard.
- Clear signals for API callers. Non-streaming execute now returns
status: "paused"with apausedReasonwhen a tool is awaiting approval, so JSON integrations can surface and resolve the request. - Approvals that wait. Pauses are backed by durable flow execution, so an approval can sit for days and survive deploys while it waits for a decision.
More control for you, with the same smooth experience for your users.