The pitch for AI agents is autonomy: set them loose and they get the work done while you do something else. The disasters come from taking that pitch literally.

The stories are always the same shape. An agent with database access "cleans up" by deleting the wrong records. A support agent issues refunds it was never meant to approve. A coding agent force-pushes over a colleague's work. None of these were model failures — the model did exactly what an unsupervised system will eventually do. They were design failures: nobody decided, in advance, which actions a machine should never take alone.

Human-in-the-loop is the fix. But most teams get it wrong in one of two directions — either no human anywhere, or a human rubber-stamping everything. The real skill is knowing exactly where the human belongs.

Autonomy is a dial, not a switch

The mistake is treating "autonomous" as yes or no. It isn't. Think of it as a dial from 1 to 5:

The point isn't to pick one number for your whole system. It's to assign a level per action, based on what that action can actually do.

The two questions that decide where the human goes

For any action your agent can take, ask two things.

Can it be undone? Reading data, drafting text, running a search — reversible, low-risk, let it run. Deleting records, sending messages, moving money, publishing — irreversible. Those want a human.

How big is the blast radius? An action that touches one test row is not the action that touches every customer. The wider the consequences, the higher up the dial you go.

Reversible and small: full autonomy, no checkpoint. Irreversible and wide: a human approves before it happens, every time. Most actions sit somewhere between, and those two questions tell you where.

Where the human actually belongs

In practice, you put a checkpoint:

Doing it without killing the point

Here's the trap: if a human has to approve everything, you haven't built automation — you've built a slow, expensive human with extra steps. Human-in-the-loop only works if the human's attention is rare and well-aimed.

A few rules that keep it useful:

Where this fits

Knowing which actions need a human — and proving your system actually enforces it — is one of the things worth auditing before you trust an agent in production. The MIMIR AI Systems Audit Checklist scores exactly this: whether high-stakes actions require approval, whether your review rate is sane, whether there's a real escalation path. It's part of the same picture as the other ways agents fail in production. €15, run it before your next deploy.

Autonomy isn't the goal. Trustworthy autonomy is — and that's the kind with a human in exactly the right places, and nowhere else.