Human in the loop for AI agents sounds simple: let a person approve what matters. In practice it fails in two directions. If every agent action needs approval, you don't have an agent — you have a slow form. If none do, you've handed over authority nobody agreed to give. The useful middle is a short, deliberate list of actions that pause for a human.
Not every agent action needs a person
Most agent actions are low-risk and reversible: looking up an order, drafting a reply, restarting a service. Holding them for approval adds delay without adding safety (rules can allow them automatically instead), and approvers quickly learn to click "yes" without reading. Approval fatigue is the most common way human-in-the-loop quietly stops working.
Which actions should pause for approval
A good rule of thumb: ask a person when an action is hard to reverse, above a threshold, or outside the agent's normal pattern. Typical examples:
- Refunds or payments above an amount
- Emails to customers about pricing, legal terms or account closure
- Changes to production systems or configuration
- Granting admin roles or changing account ownership
- Exporting customer data in bulk
Let the process owner set the threshold
The person who owns the process should set the line. Finance decides which refunds need a second look; security decides which role changes do. Good thresholds are concrete — an amount, a data category, an environment, an audience — so a rule can check them automatically.
Route it to a named approver
"Someone in finance" is not an approver. Name a role with a clear owner, and route the request to them in the tool they already use. Shared queues tend to become nobody's queue.
Give the approver context
An approval request should show the agent, the action, its arguments, the customer or system affected, and the rule that held it. With that, a decision takes seconds. Without it, the approver has to investigate — or guess. Here's what a good approval request looks like.
Decide what happens on silence
People go on holiday. Decide in advance how long a held action waits, who it escalates to, and what happens if nobody answers. The safe default is that an unanswered action does not run, and the agent tells the user it's being reviewed.
Record every decision
Every approval or rejection should be recorded with the approver, the time, and the rule that held the action. That record is what lets you answer "who allowed this?" later — see what to log for every agent action.
Go deeper
- AI agent approval workflow — routing, context, timeouts and escalation
- Human on the loop vs human in the loop — choosing the right oversight
- LangGraph human in the loop — interrupts and approvals in LangGraph
- Excessive agency — why autonomy needs limits
Where Gaten fits
Gaten holds the actions your rules say need a person, routes them to the named approver with full context, escalates on timeout, and records the outcome. Everything else runs or is denied automatically. See how approvals work.
Frequently asked questions
What does human in the loop mean for AI agents?
It means certain agent actions don't run until a named person approves them. The agent pauses, a human decides, and the decision is recorded.
Doesn't human approval defeat the point of an agent?
Only if it's applied to everything. When it's limited to high-risk actions, the agent still handles the vast majority of work on its own.
How fast should approvals be?
Set a time limit per action type — minutes for production incidents, hours for supplier payments — with escalation if nobody answers.

