Human-in-the-Loop Design

There are two ways to get human-in-the-loop design wrong. The first is to leave humans out: the agent sends the email, issues the refund, or deploys the change, and you find out afterward. The second is quieter but just as dangerous: the agent asks for approval on everything, the approvals pile up, and within a week people are clicking "Approve" without reading. An approval nobody reads is worse than none, because it gives you false confidence.

Good human-in-the-loop design puts people exactly where their judgment matters and nowhere else. This lesson shows you how to decide which actions need a human, how to design approval requests people can actually evaluate, how to implement an approval gate in your own agent loop, and how to keep approvals rare enough that they stay meaningful.

Why humans still belong in the loop#

Models keep getting more capable, so it's fair to ask why approvals matter at all. A few reasons hold regardless of model quality:

• Irreversibility. Some actions can't be undone: a sent email, a payment, a deleted record without backups, a public post. A few seconds of human review is cheap insurance against permanent mistakes.
• Injection defense. As covered in the prompt injection lesson, a hijacked agent can be talked into almost anything. A human reviewing the exact outbound action is one of the few defenses that doesn't depend on the model.
• Accountability. When an action has legal, financial, or reputational weight, someone needs to own the decision. "The AI did it" isn't an answer customers or regulators accept.
• Judgment calls. Some decisions depend on context the agent doesn't have: office politics, a customer's history, the tone your company wants to strike.
• Learning. Every approval and rejection is data about where your agent is reliable. You'll use that data to expand its autonomy safely.