Autonomy Levels
Autonomy levels are graded scales that describe how much of a task a machine performs by itself and how much responsibility remains with a person, running from full human control at one end to full machine independence at the other. The idea predates AI agents: it borrows its logic from automation research and, most visibly, from the driving-automation scale that engineers and the public already know. For AI agents, the point of a scale is to replace the vague question of whether a system is "autonomous" with a more useful one — who decides, who acts, who monitors and who can intervene at each step.
The best-known template is the SAE driving scale, SAE J3016, first published in 2014 and since revised. It defines levels 0 to 5: no automation, driver assistance, partial automation, conditional automation, high automation and full automation. Its most instructive feature is not the numbering but the handoff: at level 2 the human must supervise continuously, at level 3 the system drives but may demand that the human take over, and at levels 4 and 5 the system must manage fallback itself within or beyond a defined operating domain. The scale also hinges on the operational design domain, the conditions under which a system is meant to work, which is a reminder that autonomy is always relative to a setting, not a single property of a machine.
An older academic lineage is the levels-of-automation work of Sheridan and Verplank (1978), later developed by Parasuraman, Sheridan and Wickens (2000). Sheridan and Verplank described a ten-step continuum, from the human doing everything, through the computer offering options, suggesting one, acting if the human approves or only after a veto window, to the computer deciding and acting without telling anyone. Parasuraman and colleagues added that automation varies by function — acquiring information, analyzing it, selecting a decision and implementing an action — so one system can be highly automated at analysis yet leave the final choice with a person. That decomposition is directly relevant to agents, which often combine heavy automation of research and drafting with a human gate before the action itself.
Several proposals now adapt these ideas to AI agents, and they are not identical. One 2025 proposal (Feng and colleagues, published through the Knight First Amendment Institute) defines levels by the user's role relative to the agent — operator, collaborator, consultant, approver and observer — so that autonomy becomes a design choice made by developers rather than an automatic consequence of capability. Vendors and labs also describe informal ladders, from autocomplete to chat with tools to agents that run for hours with checkpoints, and a separate line of work such as Google DeepMind's "Levels of AGI" paper (Morris and colleagues, 2023) grades capability and generality rather than the oversight relationship. Because the terms are used loosely, a claim of "level 4 autonomy" is only meaningful if it names the scale and the task domain.
Human oversight is what the levels are really about, and it comes in distinct forms. A person can be in the loop, approving each action; on the loop, monitoring a running system and able to stop it; or out of the loop, reviewing results afterward. Which is appropriate depends on the cost and reversibility of mistakes, the speed at which the system acts, and how reliably a human can actually catch errors. Oversight also fails in known ways: approval fatigue, where constant prompts lead to reflexive clicking, and automation bias, where people trust a confident output and stop checking. The driving analogy makes this vivid, since a person asked to monitor a system that almost never needs help is poor at staying alert.
The analogy has limits, and misusing it is a common pitfall. A road has defined rules, standard sensors and measurable failures per mile; an agent operating across email, code, money and the open web has open-ended actions, no single operating domain and harms that are harder to count. Autonomy is also not one dial: an agent can have wide tool access but narrow decision rights, or narrow tools but no review, and the safest design sets permissions per action type and escalates on uncertainty or irreversibility. Higher autonomy is not inherently better either; for many workflows the right level is the lowest one that delivers the benefit, with a person approving the steps that cannot be undone.
Read autonomy levels as a vocabulary for dividing the work between machine and person: borrowed from driving and from decades of automation research, it asks for each function who decides, who acts and who can step in, and its value lies in choosing the lowest level of independence that still delivers the benefit rather than in climbing the ladder for its own sake.