AI systems / 10 Jun 2026

Why agents need approval layers

The pitch for AI agents is usually full autonomy. Set it loose, remove the human, let it run. I think that is backwards. Autonomy is not the goal. Output you can trust is the goal, and the fastest route to trusted output is an agent that knows when to stop and hand the step back to a human.

What an approval layer is

An approval layer is a deliberate gate. The agent does the work, then stops at a defined point and waits for a person to say yes.

Every agent I run inside Evolve Skateboards has one. They stop in different places:

The blog agent writes a full regional article, builds an approval pack and posts it to Slack. Nothing reaches the store until someone approves it.
The Shopify agent can prepare a product update, but it shows the exact change first and waits.
The social agent answers public comments through layered rules, and anything risky gets routed away from the model entirely.
The same agent can hide ad comments, but that runs in dry-run mode. It records what it would have done, and a human reviews the record.

None of this is a limitation we tolerate. It is the design.

Mistakes are not equally priced

The reason is economics, not caution. The cost of an error depends entirely on which side of the gate it happens.

A wrong draft costs a click. Someone reads it, rejects it, the agent goes again.
A wrong public reply costs trust.
A wrong price on a live product costs money and cleanup.

The agent does most of the work. The human does the judgement. That split is what makes the whole thing usable.

Approval makes teams faster, not slower

This is the part people get wrong. A gate sounds like friction. In practice it is the opposite, because reviewing is faster than creating.

A blog post that took the team zero minutes to write takes two minutes to approve. A product change that used to mean logging into a dashboard, finding the record and editing it carefully becomes reading one Slack message and replying yes.

The gate is also why the team actually uses the agents. Nobody hands real work to a system they cannot check. Give people a clear view of what the agent is about to do and they relax, delegate more and move quicker. Trust is the throughput.

Autonomy is earned

The gates are not permanent. They move.

A new capability starts in dry-run. It records its decisions and a human reads them. When the decisions are consistently right, the gate moves back a step. The comment agent earned live replies this way. The ad-comment hiding has not earned it yet, so it still runs dry.

This is the same way you treat a new hire. You do not give someone the keys on day one because they interviewed well. They earn each permission by being right repeatedly while someone is watching.

Map your own gates

If you are building an agent, the useful question is not "can it do this on its own." It is "what does a mistake here cost, and who should be standing at the gate." Here is a prompt to work that out for any workflow you are about to automate.

Prompt you can use

You are helping me design an approval layer for an AI agent.

The workflow I want to automate is:
[describe the workflow]

Walk me through it step by step. For each step:
1. What is the worst realistic mistake the agent could make here?
2. What does that mistake cost — a click, money, trust, a legal problem?
3. Is this step "preparation" (reading, drafting, searching) or "commitment" (sending, publishing, changing live data)?

Then recommend exactly one place to put a human approval gate: the last cheap step before the first expensive one. Tell me what the human should see at that gate to approve in seconds, and what the agent should do while it waits.

The rule

If your agent has to be perfect to be useful, the design is wrong.

Build the gate where mistakes get expensive. Then let the agent run hard up to it.