Autonomy Level

There's no single official standard that defines how many 'levels' of AI agent autonomy exist or exactly what separates one from the next — unlike, say, the well-known levels used for self-driving car automation. What exists instead is a loose, commonly used way of describing roughly how much an agent does on its own: something like manual (a human does the work, the AI at most suggests) through assisted, supervised, checkpointed, and fully autonomous, with real systems landing at different points depending on the action in question.

Because it's informal, different teams and products draw the lines differently — one vendor's 'supervised' might look like another's 'checkpointed.' The scale is still useful as a shared vocabulary for describing roughly where a system sits and for comparing two systems' design choices, as long as it's treated as a rough, descriptive framing rather than a certification an agent passes or fails.

Autonomy level often isn't a single fixed setting for a whole system anyway — the same agent might be nearly fully autonomous for low-stakes actions (reading files, running tests) while sitting much further toward the manual/checkpointed end for high-stakes ones (spending money, deleting data), which is part of why human-in-the-loop checkpoints are usually scoped to specific action types rather than applied uniformly.

How it works

The diagram lays out five informal points along the spectrum — manual, assisted, supervised, checkpoints, full autonomy — moving from a human doing the work with the AI merely suggesting, through stages where the AI does more of the acting but still pauses at defined points, to an end where it acts without pausing at all. There's no highlighted 'correct' point on this diagram deliberately — where a given deployment should sit depends entirely on the risk of the actions involved, not on a target every agent should aim for.

Example

A note-taking agent might sit near 'full autonomy' for organizing and tagging notes (low stakes, easy to undo), while an agent with access to a production database might be deliberately kept at 'checkpoints' — acting mostly on its own but always pausing before a destructive query — even if it's technically capable of running unsupervised.

How it differs

Common misconceptions

FAQ