What is human in the loop?
Human in the loop (HITL) is a design approach in which a person reviews, approves or corrects an automated system's work at defined points, so that decisions and actions with real consequences pass through human judgment before they take effect.
How does human in the loop work with AI agents?
Human in the loop puts a person at the points where a mistake would cost the most. The term comes from control systems and machine learning, where people labeled examples and corrected model output. With AI agents that take actions, the question becomes: which actions should wait for a person?
There are three places to put a checkpoint:
- Before an action. The agent proposes what it will do and waits for a yes. This is the most common pattern for anything that spends money, deletes data or contacts a customer.
- During a task. The agent stops to ask when something is ambiguous ("Two accounts match that name. Which one?") instead of guessing.
- After the fact. People review completed work, spot-check a sample, or read an audit log, and can undo or correct it.
A related term, human on the loop, describes a person who monitors and can step in but doesn't approve each action.
Which agent actions need human approval?
Sort actions by consequence, not by how complex they are. Four questions help:
- Can it be undone? Renaming a project can. Sending an email to a customer can't.
- How much does it touch? One record is different from every record in the account.
- Does it involve money, access or legal terms? Refunds, permission changes and contract edits almost always need a person.
- Is the request clear? A vague request deserves a clarifying question before any action.
Here is how that plays out in a billing tool. An admin asks the agent to refund the last three invoices for a customer. The agent finds the invoices, totals them, and shows a confirmation card with the amounts and the customer name. Only after the admin confirms does it issue the refunds, then it reports the result with links to each one.
How do you add human approval without slowing everything down?
Approve the risky few, not the routine many. If every step needs a click, people approve without reading and the checkpoint stops protecting anyone.
Three habits keep it useful: make the approval screen show exactly what will change, let low-risk actions run without asking, and run every action with the signed-in user's own permissions so the agent can never do more than that person could. Before launch, test the agent against real requests so the approval points sit where the errors actually happen.
Foldspace's Product Agent has these controls built in: it can't act beyond the signed-in user's permissions, a team can add a confirmation message to any action so the user approves it before it runs, and no action reaches users until the team publishes it.