Where should it stop? Teaching children to design approval gates
Autonomy is a dial, not a switch. Here is how we teach a child to decide exactly which steps an agent may take alone, and which ones must stop and wait for a human to say yes.
A twelve-year-old built an agent to help finish her science fair project - checking progress each evening, drafting the next step, and, in her first version, emailing her teacher directly whenever a step was complete. We asked her to picture the agent making a mistake on step three. She thought about it for a second and then rewrote the whole thing so that nothing left the house until she had personally looked at it and said yes.
That rewrite, done entirely by an eleven-year-old with no prompting beyond one question, is the whole idea behind an approval gate: a point in an agent's steps where it must stop and wait for a human, no matter how confident it is that it's right.
What is a human-in-the-loop gate?
It is a deliberate pause built into an agent's chain of steps - a place where, instead of continuing automatically, the agent stops and shows a person what it's about to do, and waits for a yes before going further. The agent can still do everything up to that point on its own. It just cannot cross the gate without permission.
Where should it stop?
Not every step deserves a gate - that would just turn an agent back into a chatbot with extra clicks. The useful question is which specific steps carry a cost if they go wrong that a person would want to catch before it happens, not after.
- Checking progress against a to-do list - low stakes, fine to run alone.
- Drafting the next day's step - still just a draft, fine to run alone.
- Sending anything to a teacher, a stranger, or anyone outside the family - a gate belongs here, every time.
- Spending money, however small - a gate belongs here too.
Autonomy is a dial, not a switch
The instinct many children start with is all-or-nothing: either the agent does everything by itself, or a person has to approve every single step, which defeats the point of having an agent at all. Neither extreme is right. The actual skill is placing gates only where a mistake would be expensive or hard to undo, and leaving everything else to run.
The honest limit is that a gate only helps if a person actually looks carefully when it fires, rather than reflexively approving everything an agent shows them. A gate that gets rubber-stamped every time provides the feeling of oversight without the substance of it - which is worth remembering before the next article, where children are asked to deliberately imagine their own agent going wrong, on purpose, before it ever runs for real.
Questions we get asked
What is a human-in-the-loop gate?
It is a deliberate stopping point built into an AI agent's sequence of steps, where the agent must pause, show a person what it is about to do, and wait for explicit approval before continuing - rather than completing every step automatically without anyone checking in between.
Which steps in an AI agent need a human approval gate?
As a rule of thumb, any step that is expensive, hard to undo, or affects someone outside the immediate task - sending a message to another person, spending money, sharing a file externally - deserves a gate. Low-stakes internal steps, like checking a to-do list or drafting a note nobody sees yet, generally do not.
Should an AI agent run completely on its own or need approval for everything?
Neither extreme works well. Full autonomy risks a small mistake compounding before anyone notices; approving every single step removes the benefit of having an agent at all. The better approach is placing gates only at the specific steps where a mistake would be costly or difficult to reverse, and letting everything else run unattended.