The short answer
This pattern is often called human in the loop. An agent can read and draft on its own, but a person decides before it sends, changes, pays or deletes something.
Why approval matters
- Language models can make mistakes that sound confident.
- Text in emails and documents can try to mislead an agent, which is called prompt injection.
- Some actions cannot easily be undone once they reach a customer or another system.
Which actions need approval
| Action | Suggested rule |
|---|---|
| Searching and reading approved data | Can usually run alone, with least-privilege access |
| Drafting a reply or report | Can run alone if nothing is sent |
| Sending a message outside the company | Ask a person first |
| Changing records in another system | Ask a person first |
| Payments, deletions and legal or HR decisions | A person decides |
How to design a good approval step
- Show the exact action and recipient.
- Choose someone with the authority to decide.
- Set a time limit, and treat no answer as declined.
- Record who decided and when.
- Check how often people change or reject drafts.
What approval cannot do
Approval reduces risk, but it does not remove it. People can approve too quickly, and anything an agent can read can still reach its model. Limit what agents can read and use as well.
How approvals work in Hivelock
In chat, connector actions that change data wait for a person, and no answer within 15 minutes counts as declined. An agent step in a workflow cannot use tools that need approval, because nobody is watching it. A connector step runs as built, so you add an approval step before it. Read the user guide.
Common questions
Is human in the loop the same as human approval?
Human in the loop is the wider idea of keeping people involved. Approval is one common form of it.
Does every agent action need approval?
Usually not. Reserve approval for actions that change things, so people can focus on the decisions that matter.
Tell us which process you want to automate. Get early access.