How does an approval work?
- You pick the tools
Tool policy lists which tools ask first. You can simulate a policy change before you publish it.
- The run pauses
When the agent calls a gated tool, the run stops before the call. The person who started it sees "waiting for approval".
- The approver sees the whole request
The approvals inbox shows the tool, its arguments, the agent and the policy that required the approval.
- The decision is recorded
Approve and the run resumes and finishes. Reject and the tool does not run. Either way the decision goes into the audit log under the approver's identity.
Why gate tools and not whole agents?
Most of what an agent does is reading and drafting, and asking a person each time would bury them. The step that changes something outside the conversation, a refund or a deleted record, is the one a reviewer wants to see. Gating at the tool keeps people on the decisions that carry weight.
What does the stop button do?
An admin presses Halt agents, types a reason and the word HALT, and confirms. New runs are refused, and runs already in progress stop at their next step boundary. Users see an organization-wide emergency stop with the reason the admin gave. Resume agents puts everything back.
Limits: the switch fails open if its own store errors, and it does not cover document-intake submissions. Approvals cover tool calls, not what an agent writes in a chat reply.