Human-in-the-Loop (v1)

You can now require your approval before an agent runs specific tools. In the agent builder, mark any tool — a native tool, a connector operation, an MCP tool, or a custom action — as "Needs Approval" instead of running automatically. When the agent goes to use that tool, the conversation pauses and shows an approval card with a plain-language summary of what it's about to do, and you choose Allow or Deny. Allow lets the agent continue; Deny stops the action and hands the conversation back to you.

Conversations waiting on you are easy to spot: a yellow indicator appears on the conversation list and in the agent library (with a "Needs approval" filter), and scheduled runs that are paused show an "Awaiting approval" status. Approvals also work inside sub-agents — a request raised by a sub-agent surfaces in your main conversation. Tools are set to run automatically by default, so nothing changes until you choose to gate a tool.