Agentic patterns · 04 of 04

Agentic patterns: Human approval gate

Before an agent takes an action that is costly or hard to undo, it pauses, shows its reasoning and waits for a person to approve, reject or redirect it.

The principle

Human approval gate

Agents are useful because they act: they roll back a deployment, pay an invoice, send a letter. The same ability makes a wrong decision expensive. An approval gate keeps the agent’s speed for the investigation and puts a person at the moment that matters, before the action.

Doing this properly takes more than a confirmation button. The run must save its state at the pause, survive a restart while it waits, and resume with the person’s answer, including a reason when they say no. Because every step is checkpointed, you can also rewind to the pause and replay what would have happened under a different decision, which helps in testing and in audits.

Two cases are easy to forget. When nobody answers, the gate needs a timer and a next person to call; it should never act on silence. And when approval comes too late, after the system has moved on, you need a way back: a compensating step that undoes or offsets what was done, decided before the gate goes live.

yesno, with a reasonagentInvestigateagentProposepersonApprove?codeActcodeVerify

codemodelagentperson

The shape of the pattern. Each box can be filled by code, a model, an agent or a person; the colours show how this page fills them.

Also known as Human-in-the-loop (LangGraph interrupts) · Human-in-the-loop workflows (Microsoft Agent Framework) · Factor 7: contact humans with tool calls (12-factor agents)

Where it applies

  • Payments, write-offs and journal entries above a threshold
  • Production changes, access grants and anything else on a change-approval list
  • Outbound letters or emails with legal or financial consequences
  • Early deployments, before you trust the agent with less oversight

Where it does not

  • Low-risk actions that are cheap to reverse, where approval only adds delay
  • Volumes so high that reviewers start approving without reading
  • Decisions the reviewer cannot judge from what the agent shows them

Fill the boxes

One shape, three ways to build it

The pattern exists with or without AI. The design decision is which boxes deserve a model, and what that costs in money, speed and auditability.

All code

A rule approves anything under a limit, as banks do with small payments. It is instant and exactly as good as the limit you set.

Hybrid, as shown here

An agent investigates and proposes, and a named person approves before anything changes.

All agents

A second agent approves the first one’s proposal. That scales, but two models can share one blind spot, and an auditor will still ask which person was accountable.

See it work

The 3am incident

03:12, Saturday. Checkout is failing. The on-call agent, Claude, investigates and proposes one production change at a time; you hold the button. Approve everything and watch it keep going after the obvious fix. Then rewind to the first pause and answer differently: reject with a hint, reject without one, or do not answer at all.

Recorded on 19 November 2026 with Claude Opus 5.5: the model's answers are real, replayed here. The production system is a small simulation with a hidden root cause.

Watch for. Pause for approval with interrupt(), resume with Command(resume=...), then rewind to the pause and fork a different future.

Building it

This demo runs on LangGraph

LangGraph runs an agent as an explicit state graph that saves every step, so it can pause for a human, resume, and rewind. Other frameworks can carry the same pattern; this is when I would pick this one.

Reach for it when

  • The process has steps you can draw, and some of them need a human sign-off
  • Runs are long or risky and must survive a pause, a crash or a redeploy
  • You need to audit or replay exactly what the agent saw at each step

Skip it when

  • One model call with a couple of tools does the job
  • Nobody will ever need to pause, inspect or rewind a run
  • Your team wants a hosted no-code canvas rather than Python
Framework
langgraph 1.2.21
Licence
MIT
Model
claude-opus-5-5
Watch time
~2 min

Advice

Questions I ask before using it

  • Which actions need a gate, and which can simply be undone instead?
  • What does the reviewer need to see to decide in under a minute?
  • What happens when nobody answers?
  • Does a rejection give the agent something to work with, or does the run just stop?

Contact

Questions, ideas or a conversation?

I’m always glad to hear from people working on AI, product strategy, transformation or software.