Your multi-agent procurement workflow can draft purchase orders and call a create-PO tool. Finance requires that any PO above $10,000 pause for human approval before the tool runs, with the ability to override or reject. How should you implement this control?
Select an answer to reveal the explanation.
Short Explanation
You need human-in-the-loop approval gates before high-risk tool invocation, with override and rejection paths and durable workflow state while waiting. That pauses the create-PO tool above $10,000 and resumes only when finance approves, rejects, or overrides. Temperature tweaks, after-the-fact emails, or removing tools don't give you a real preventive control.
Full Explanation
The correct control is human-in-the-loop approval gates before high-risk tool invocation, with override and rejection paths and durable workflow state while waiting. Production multi-agent systems combine autonomy with HITL for high-impact actions. Implement an approval node (Agent Framework / workflow engine) that intercepts tool calls above policy limits, notifies approvers, stores pending state, and only then allows the create-PO tool to execute. Support explicit reject and human override paths for exceptions. Pair HITL with audit logs, correlation IDs, and clear agent personas that know when they must escalate. Raising the agent temperature only for amounts above $10,000 is incorrect: sampling temperature is unrelated to authorization or approval semantics, and temperature tweaks do not create approvals. Logging the PO after creation and emailing finance with no blocking step is incorrect: post-creation notification is detective, not preventive, and fails the “before the tool runs” requirement. Disabling all tools and having the agent only return markdown recommendations forever is incorrect: permanently removing tools over-constrains low-risk automation, abandons automation below the threshold, and is not a selective HITL design.