Essay · Draft
Rate the Action, Not the Agent
The debate about how much autonomy to give AI agents is asking the wrong question. Every till in the world already has a better one.
Kamel Dhakwani2026-02-24 · 1 min read
Asking how much autonomy to give an AI agent is the wrong question. It has no stable answer, because it changes with every model release, and it turns governance into an argument about trust.
Point-of-sale software settled this a long time ago. The same cashier rings up sales freely and is then stopped on the next screen by a prompt for manager authorisation. Nobody takes this as an insult. The till is not rating the cashier. It is rating the action.
Agents should work the same way. Every action they can take carries a level:
- Green. The agent acts.
- Orange. The agent prepares the work and asks.
- Red. A named person decides.
Two details matter more than the colours.
The first is that the level belongs to the action in its context, and never to the model. A better model does not make a refund less sensitive.
The second is where the rule lives. A cashier cannot talk their way past the manager prompt, because the prompt is not part of the cashier's training. It is in the software. If your approval rule lives in the agent's instructions, it is a polite request. If it lives in the code that runs when the tool is called, it is a control.
The objection
The strongest argument against this is approval fatigue. Ask people to confirm everything and they stop reading, and the control becomes theatre. A till handles this by staying out of the way: the ordinary sale goes straight through, and only the exceptions stop. An approval system earns its place the same way, by being rare enough that people still look.
Ask which actions you can afford to get wrong, and the autonomy question answers itself.