Human in the loop (HITL) means a person approves an AI agent’s consequential actions before they happen. It turns “the AI did something” into “I decided, the AI did the work”.
What it means
In a human-in-the-loop system the agent works on its own through routine steps and pauses at checkpoints for a person to approve, correct or cancel. The human does not do the work; the human keeps the decision.
When a checkpoint matters
Not every click needs approval — that would make an agent slower than doing it yourself. The useful rule is reversibility. Scrolling, opening a page or typing into a field can be undone. Sending a message, submitting a form, booking, paying, subscribing and deleting cannot. Those are where a checkpoint belongs.
The autonomy spectrum
- Ask every step — safest, slowest.
- Ask at final steps — routine work is automatic, consequences wait for you. This is Hark’s default and cannot be turned off.
- Autonomous with automated checks — the agent runs freely while classifiers watch for problems; faster, with more trust placed in the system.
Different products choose different points. See Hark vs Claude in Chrome for a concrete contrast.
How Hark does it
Before any step that sends, submits, books, orders, pays, subscribes or deletes, Hark shows a card: “Your approval needed — Hark wants to: Click ‘Book now’”, with Approve and Cancel. Logins, verification codes and payment details are handed to you entirely. You can press Stop at any moment.
What a good approval prompt looks like
It names the exact action in plain words, appears before the action and not after, defaults to doing nothing, and is rare enough that you still read it.