Human-in-the-loop

#How human-in-the-loop works

The agent names the action it wants to take. A gate holds the action until a person answers. The gate shows what will happen, with the exact data that will be sent. If the person says no, or does not answer, the action does not run. An approval can cover one action, one run or a whole session. The narrower the approval, the less harm a mistake can do.

#Why it matters for AI agents

Reading a page carries little risk. Changing something on a site carries more. An agent that can book, buy or message for the user needs a stop before its first write. Human-in-the-loop does not stop the agent from reading or planning. It stops the step where something changes. Pair it with least privilege so the agent can reach only what it needs, and with an audit log so every action can be reviewed later.

#Wield and human-in-the-loop

Wield applies this to every write. A recipe that changes data, such as a booking, a message or an order, stops before its first write request. A window in the user's browser shows the site, the task, and every request with its body. Tokens in the body are masked. Deny is the default. Two minutes without an answer also counts as a deny. One Allow covers one run. No setting, policy or agent can turn this off, and an organization policy can only restrict what runs.

#Related terms

• Prompt injection: the attack that human approval helps to stop from becoming a real action.
• Least privilege: the limit on what an agent can reach in the first place.
• Audit log: the record of what the agent did, for review after the fact.

#FAQ

#What does Wield ask before booking something?

Wield opens an Allow window before the booking's first write request leaves the browser. The window lists each request with its method, its path and the body as it will be sent. You can allow the run or deny it. If you do nothing for two minutes, the request is denied.

#Can I turn off approvals?

No. Wield's approval for writes cannot be switched off. No setting, admin policy or agent can skip the Allow window. An organization can restrict which sites and actions are allowed, but nothing can widen what runs without the user's Allow.