Human approval for AI agents, only for the actions that need it.
Your policy marks which agent actions are sensitive. Each one waits for a named person, while the agent's other calls keep running.
What needs approval
Some agent actions need approval.
Your policy names which ones.
With Hawcx, human approval for AI agents means each action your policy defines as sensitive, such as a production rollback, a large refund or a contract, waits for a named person. They see the exact call and why it was held. Every other call within the agent's ceiling runs.
A store cashier can ring up sales all day, but to void a sale or open the drawer, a manager has to turn a key.
While it waits
Held means waiting, not failing.
The agent's other work runs on.
A held action waits for its approver while the agent's other calls within its ceiling keep running, so one approval doesn't stall the whole task. The hold is kept by the Hawcx Authorizer, outside the model. The model can be tricked. Enforcement does not depend on it.
AI agent authorization on every tool callWhat the approver sees
The approver sees the whole request.
And the reason it was held.
Each held request shows its approver four things. Then they approve, edit, reject or hold it.
- The agent
- Which agent is asking, by its Hawcx identity.
- Who it acts for
- The employee the agent works for.
- The action
- The tool call, exactly as the agent made it.
- The reason
- Why your policy marked it sensitive.
Each approval is recorded on the action's signed receipt.
AI agent audit trailApproval questions, answered.
Before this: AI agent authorization on every tool callAfter this: AI agent audit trail
Which AI agent actions need human approval?
The AI agent actions that need human approval are the ones your policy marks as sensitive, such as a production deploy, a payment over a set limit or sending a contract. Hawcx holds each one for a named person. Anything else the agent does within its ceiling runs without waiting.
What does the approver see?
The approver sees the whole request: which agent is asking, the employee it acts for, the tool call exactly as the agent made it, and the reason your policy held it. With that in front of them, they approve, edit, reject or hold the request. Each approval is recorded on the action's signed receipt.
What happens to the agent while an action is held?
While an action is held, only that action waits: the agent carries on with its other calls, as long as each stays within its ceiling. The held action runs once it is approved. The Hawcx Authorizer keeps the hold outside the model, so the agent can't release it on its own.
Who approves a held AI agent action?
A named person approves each held AI agent action: whoever your policy routes it to. Your policy can route by action or by agent, to the employee the agent works for or to someone else. The decision stays with that person, never with the agent or the model steering it. For rollout across many teams, see Hawcx for Enterprise.
Can the approver change a request before approving it?
Yes. The approver can edit a held request, or approve, reject or keep holding it. They decide with the full request in front of them, including why your policy held it. The held action waits on their decision.