Autonomy Slider
The autonomy slider is Andrej Karpathy‘s name for a control, left in the user’s hands, over how much of a task an AI system does without supervision. He set it out in his 2025 talk at YC’s AI Startup School as the defining feature of a good partial-autonomy product.
The point of the image is that autonomy is a setting, not a switch. At one end the human does the work and the model assists; at the other the model acts alone and the human reviews the result. Karpathy’s argument is that fallible models call for products that let the user choose a point along that range task by task — and that are designed so the setting can move towards full autonomy as the models improve.
It is easily confused with agents in general. An agent is a system that acts on its own; the slider is a claim about product design — that the degree of agency should be adjustable and visible, rather than fixed at ‘fully autonomous’ because demos look better that way.
Examples from the talk
- Cursor, the AI code editor, offers tab completion (the human in charge), an edit to a selected chunk, an edit to a whole file, and a mode that works across the entire repository — four positions on one slider.
- Perplexity offers a quick search, a research mode and a ‘deep research’ run that returns ten minutes later.
- Tesla Autopilot, where Karpathy worked for five years, showed the driver what the neural network saw and took on more of the driving over time.
Why a slider rather than full autonomy
The human is still the verifier, so the product’s job is to keep the loop between the model’s generation and the human’s verification fast. Two things help: interfaces that let people check work by looking rather than reading (a diff in red and green), and keeping the model ‘on the leash’ with small, concrete increments. Karpathy’s scepticism about full autonomy rests on driving: his first self-driving ride, in 2013, had zero interventions, and twelve years later the problem is still not solved. Hence his summary image: build ‘Iron Man suits’ — augmentation the wearer controls, able to fly alone when needed — rather than ‘Iron Man robots’.
In the wiki
- Andrej Karpathy on Software 3.0, LLMs as Operating Systems, and Partial Autonomy — the talk that introduces the term
- Andrej Karpathy — who coined it
- Agentic Engineering — the professional practice of working with agents at the high end of the slider
- Vibe Coding — the casual end of the same range