Human-in-the-Loop: Where People Should Still Check the Work

Full autonomy sounds efficient until an agent does the wrong irreversible thing. Human-in-the-loop means choosing checkpoints on purpose, not everywhere and not nowhere.

The core idea

Human-in-the-loop (HITL) means a system pauses at certain points and waits for a person to approve, edit, or reject what it's about to do, instead of running start to finish alone. It's a design choice about where to put a checkpoint, not a statement that automation can't be trusted.

Where checkpoints tend to make sense

Irreversible actions - deleting data, messaging someone outside your team, spending money - are the classic case: once it happens, there's no undo. Low-confidence situations are another: if a task looks unlike anything handled before, that's a good moment to ask a person. High-stakes but rare decisions, like a production deployment, are worth a checkpoint even if the system is usually right.

What a checkpoint looks like in an AI coding agent

Often this is simply an agent proposing a specific action - deleting a file, running a migration, pushing to production - and pausing until you approve it. Some systems let you approve as-is, edit the action, or reject it with feedback the agent can act on.

The risk of auto-accepting everything

Turning off all checkpoints removes friction, but also removes the last chance to catch a mistake before it becomes real - a wrong file deleted, a wrong deploy pushed. The right amount of human-in-the-loop isn't zero and isn't everything; it's putting the pause exactly where a mistake would actually hurt.

EXAMPLE

A concrete rule: any tool call that writes outside the current project folder, runs a database migration, or sends an external message pauses for approval. Reading files and running tests do not.

🛠️ EXERCISE — TRY IT YOURSELF

Design a checkpoint rule for an AI agent working on your own project.

List 3-5 actions an agent might take in your project, from harmless (reading a file) to risky (deploying, deleting).
For each one, decide: reversible or not, and how costly a mistake would be.
Mark which ones should require approval and which should run automatically.
Write the rule down in one or two sentences, the way you would configure it in a real tool.

✅ SELF-CHECK

☐ Did you note for each action whether it's reversible or irreversible?
☐ Does every checkpoint sit before execution, not after?
☐ Is every checkpoint technically enforced, instead of just a request in the prompt?

QUICK QUIZ

What's the key design mistake to avoid when adding human-in-the-loop checkpoints to an agent?