AgentKey · Controls

Human Approval for AI Agents

Human approval for AI agents (human-in-the-loop) is a control that pauses a risky agent action until a person explicitly approves or denies it, turning an autonomous decision into a supervised one.

By AgentKey · Last updated 2026-09-01

When to require approval

Not every action needs a person. Approval is for actions that are irreversible, high-impact, or outside the agent's normal lane: issuing a large refund, sending email outside your domain, deleting records. The policy decides which actions ask; risk decides when an otherwise-allowed action still escalates.

The approval flow

Authorization returns an ask decision instead of allow or deny.
A pending approval request is created with a reason and an expiry.
A person approves or denies it, on the dashboard or by API.
The action proceeds only after approval; a denial or expiry blocks it.
The resolution is recorded on the evidence chain.

Expiry and latency

An approval that waits forever is an approval that stalls the agent. A time-to-live lets a request expire and fail safely, so an unattended ask becomes a block rather than an open door.

Where AgentKey fits

AgentKey's ask decision creates a pending approval request with a configurable TTL. resolve_approval approves or denies it; expired approvals are failed safely. Every resolution is written to the evidence chain.

Frequently asked questions

How can AI-agent tool calls require human approval?

Set the permission decision to ask, or let risk escalate an allow into an ask. The action is paused, a pending request with a reason and expiry is created, and a person approves or denies it before the call proceeds.

What happens to an approval nobody responds to?

It expires and fails safely, so the action is blocked rather than left open indefinitely.