The Autonomy Slider

Give AI as much autonomy as you can verify quickly — and let people move the slider up as trust grows.

More autonomy means bigger changes to verify, so set the slider where checking stays fast and raise it as trust grows.

To an executive

AI tools work best when people choose how much to hand over, from a small suggestion to a whole task. That keeps risk proportionate while trust builds.

To an engineer

Offer a range from suggestion to autonomous action, and keep each change small enough to verify quickly. Checking AI output, not generating it, is the bottleneck.

The idea

In a 2025 keynote at Y Combinator's AI Startup School, Andrej Karpathy argued that the useful AI products of the near term are partially autonomous. Rather than aiming for full automation, they let people choose how much to hand over, from a small suggestion to a large, independent change. He called this an autonomy slider, pointing at coding tools that range from completing a line, to editing a selection, to changing whole files, to acting as an agent.

The design goal is a fast loop between the AI generating work and a person verifying it. The more autonomy, the bigger each change — and the harder it is to check.

Why it works

Verification is the bottleneck. An AI that produces a thousand-line change in seconds saves no time if checking it takes hours.

Trust grows unevenly. People are happy to hand over some tasks long before others; a slider lets them do that.

Where it applies

Designing AI features in products and internal tools.
Planning automation roadmaps, moving step by step from assistance to autonomy.
Risk decisions about which actions an AI may take without approval.

How to apply

Define levels. For each task, list what the AI can do at low, medium and high autonomy.
Make verification fast. Show diffs, sources and summaries so people can check work at a glance.
Default low for risky actions, and let users raise the slider where outcomes are easy to check or undo.
Measure acceptance. How often people accept, edit or reject AI work tells you where more autonomy is earned.

Example scenario

A finance team wanted AI to reconcile accounts end to end. Instead, the first release suggested matches that a person confirmed with one click, showing the evidence for each.

After three months with very high acceptance on simple matches, those were automated, while complex matches stayed at suggestion level.

Questions to ask

How fast can a person check what the AI did at each level of autonomy?
Which actions are safe to automate fully, and which need approval?
Can users choose how much autonomy to give, task by task?

When not to use it

Fully deterministic tasks, where ordinary automation is simpler and more reliable.
Actions with no safe partial step, where the right answer is either a person or a tested system, not a slider.

How it misleads you

A slider can push the burden onto users. If every decision about autonomy is left to the person, they will either approve everything or turn the AI off. Set sensible defaults and limits by design.