The approve button is where human-in-the-loop quietly fails

The most common way a human-in-the-loop agent fails isn't dramatic. Nothing crashes. Nobody gets paged. It's the approve button.

In week one, people read every draft the agent produces. They fix the tone, catch the wrong customer name, push back on a made-up number. By week six, "approve all" is muscle memory. The gate is still there on paper. It just stopped doing anything.

Our playbook calls this a soft gate, and it's the first failure mode I look for in any Phase 2 (Gated Agents) rollout.

It shows up in the data, not just in anecdotes

A 2026 preprint, "Habituation at the Gate", followed 400 repeat reviewers of pull requests written by AI coding agents in open-source repos. Comparing each reviewer's early reviews with their later ones, approval rose from 30.1% to 36.8%, and inline comments per review fell 22%. Reviews also sat in the queue longer, so people weren't reading more carefully. They were reading later and saying less.

Two honest caveats. The authors say they can't establish causality, since the agents also got better during the study. And it covers open-source repos only, not your company's approval queue. But the pattern will look familiar to anyone who has run one: more approvals, less scrutiny, longer queues.

Why it happens

A gate that almost never catches anything trains people to stop looking. The agent is right 95 times, so the 96th approval gets a glance. Add a growing queue and a reviewer who also has a day job, and "approve" becomes the cheapest way to make the notification go away.

That isn't a character flaw. It's what a well-designed gate has to plan for.

The repairs are boring on purpose

None of these need a new tool. They need someone to own them.

• Draw the gate as a checklist someone could run at 2 a.m. Approve, edit, reject, escalate. If the reviewer can't say what they're checking for, they're not checking.
• Make the edit rate visible. Track how often approved outputs were edited first. If nobody ever edits, ask why. Either the agent is perfect or the review isn't happening.
• Sample-audit what got approved. Pull a handful of approved outputs each week and have a second person review them properly. Don't just trust the log that says a human clicked the button.
• Keep a golden set. Ten to fifty "correct / wrong / must never ship" examples, run before every prompt or tool change. (Friday's post goes deeper on this.)
• Script the failure drill. When the agent is wrong and it got approved anyway: who notices, who halts, who notifies, who fixes the golden examples.

The paper's authors suggest a few more that fit right alongside these: rotate reviewers so nobody owns the whole queue, flag long unbroken approval streaks for a second look, and show reviewers their own approval trend.

Not every action needs the same gate

Part of what wears reviewers down is gating everything the same way. AWS's Well-Architected Agentic AI Lens lists "applying uniform oversight regardless of risk" as an anti-pattern, and recommends sorting agent actions into autonomous, notify, and approve tiers based on impact and reversibility.

That matches how I think about it. Save the human's attention for the steps where a mistake is expensive or can't be undone. A reviewer with ten meaningful approvals a day will stay sharper than one with two hundred trivial ones.

Gated Agents isn't a waiting room

It's tempting to treat Phase 2 as a stop on the way to "real" autonomy. For anything that reaches people outside the building, it's often the right place to stay, and for some outputs, like anything that reaches a child or family, the gate stays permanently. So it's worth being excellent at.

What to do this week

Pick one approval queue you own. Pull last month's approved outputs and count how many were edited before approval. If the answer is close to zero, pull ten at random and review them yourself. That's your baseline, and it'll tell you fast whether your gate is still a gate.

Get the playbook

The Phase 2 chapter, including the gate checklist, failure drill, and the rest of the failure modes we watch for, is in the free 231 Labs AI Transformation Playbook at 231labs.com/playbook.