What is human-in-the-loop AI?

In the loop, on the loop, or out of it?

Three arrangements, routinely confused:

• Human in the loop: approval is required before the action. Slower, and the only arrangement that prevents a specific bad action.
• Human on the loop: the system acts and a human supervises, able to intervene or roll back. Faster, and it catches problems after they happen.
• Fully autonomous: nobody in the path. Appropriate for reversible, low-stakes, high-volume actions.

What does an approval gate actually look like?

A card. A companion reaches an action it treats as consequential (sending an email, writing to a system of record, deleting something) and the browser stops and shows you the recipient, the destination and the exact text or field values. You approve once, approve for the session, allow it always, or refuse, and the choice is remembered per tool rather than per conversation.

The detail that makes the gate real is the absence of fan-out. One message is one card, so "send this to thirty people" is thirty decisions. Ask Always mode in Settings → AI permissions puts every step behind a card. The wider model is on the security page.

How approval gates fail

Approval fatigue, and it is not hypothetical.

Ask a person fifty times a day and they start clicking through without reading. An unread approval is functionally auto-approve, with a false sense of safety attached.

The defenses are unglamorous. Gate few enough things that each card is worth reading, show the actual payload rather than a summary of it, and keep batch sizes small enough for a human to check. If a workflow needs a hundred approvals, redesign the workflow.

Strawberry is built for that. Reading and navigating run on their own, only consequential steps raise a card, and the card carries the actual recipient, destination and field values rather than a summary of them.

Frequently asked questions