What is earned autonomy for AI agents?

Earned autonomy

Earned autonomy means an AI agent earns the right to act on its own, one person and one action at a time, by being right again and again. It drops back to asking when something is new, big or strange.

How does earned autonomy work?

The agent climbs a ladder with each person, and keeps a separate ladder for each kind of action.

1. Suggest. The agent says what it would do. The person does it.
2. Prepare. The agent does the work and stops one click short. The person approves.
3. Act and tell. The agent acts on small, reversible things and reports back straight away.
4. Act and summarise. The agent handles a whole category and sends a digest.
5. Act and escalate. The agent runs the area and only asks when something is unusual.

It earns a rung with a track record the person can see, where mistakes are cheap, or when the person asks for it ("just do it next time"). The next rung is offered out loud, never taken silently. When something doesn't match what it has seen before, it steps itself down and asks. After a mistake it drops one rung on that ladder, not all of them.

What does earned autonomy look like in practice?

Say an agent sorts out your gig tickets. At first it only suggests. After a few yeses it fills the basket and waits for one click. A few weeks later it buys and tells you after, and by the top rung it books your usual nights without asking.

Then one Sunday the tickets you want are £340 each on a resale site. An agent that only knows its permissions would buy them, because it's allowed to. An agent with earned autonomy notices this doesn't look like the last six times, steps itself back down and asks. That message costs you five seconds. Getting it wrong would cost you £680, and your trust in everything else it does.

Why not just give the agent an autonomy setting?

A settings screen on day one gives people nothing to base the choice on. Most pick the lowest and never come back. And trust doesn't transfer. Someone who lets the agent buy £30 gig tickets without asking hasn't said anything about rent, or about emailing their boss. One switch for the whole product can't hold both.

What decides how high an action can climb?

Mostly whether it's a one-way or a two-way door. A two-way door is easy to undo, like moving a calendar event, and can climb to the top quickly. A one-way door can't be undone, like sending a gift or buying a ticket you can't refund. It needs far more evidence, and some should stop at "prepare" for good. The full rules, with tables, are in the earned autonomy post.

How is earned autonomy different from levels of autonomy?

Frameworks like Levels of Autonomy for AI Agents by Kevin Feng, David McDonald and Amy Zhang describe levels by the role the user plays, and treat the level as a design decision made before launch. Earned autonomy also has rungs, but the rung is set per person and per action, and it changes while the agent runs. It goes up with a track record and down after a mistake.

Where does the idea come from?

From how trust works with a new person on a team. In week one they suggest. In month one they act on small things and tell you after. By month six they handle whole areas and only escalate the odd case. Nobody signs a form for each step. I set it out for agents in Earned autonomy: how much should an agent do before it has your trust? in October 2026, and applied it to music in What should a music agent do without asking?.

It's also code. earned-autonomy is an open-source decision engine (Apache-2.0) that answers allow, ask or deny for an agent, a person and an action, with the reason. It uses six levels rather than five, splitting "ask first" from "heads up, doing this in ten minutes unless you stop me".

Related ideas

• One-way and two-way doors for agents. How costly an action is to undo decides how far it can climb. In the earned autonomy post.
• The Yes Machine. Why trust falls off a cliff after one confident wrong answer. In The Yes Machine.
• The agent scope doc. One page before the PRD on what an agent can see, do and must never do. In Your PRD template is wrong for agents.

FAQ

What is earned autonomy? Earned autonomy means an AI agent earns the right to act on its own, one person and one action at a time, by being right again and again. It drops back to asking when something is new, big or strange.

What happens when an agent with earned autonomy gets something wrong? It drops one rung on the ladder for that kind of action, not all of them. A single bad purchase sends ticket buying back to "prepare" for a while, without wiping out everything else it has learnt about the person.

How do you know earned autonomy is working? Measure it per action type. Acceptance rate, edit rate, undo rate, how often the agent was right to step itself down, and how often people say "just do it".

Is there open-source code for earned autonomy? Yes. earned-autonomy on GitHub is a decision engine under Apache-2.0. You give it the agent, the person, the action and the context, and it answers allow, ask or deny.

To cite this, use “Earned autonomy”, Matt Alcock, mattalcock.com/ideas/earned-autonomy.