A way of describing how much an AI system decides and does on its own, from merely suggesting to acting without anyone watching. The idea is borrowed from self-driving cars, which are graded from no automation to full automation. For AI agents, the levels describe what the person does: whether they drive every step, work alongside, approve proposed actions or simply observe. Picking a level is a deliberate choice about risk, not just a question of what the system can do.
Technical view
The best-known precedent is the J3016 taxonomy from the standards body SAE International, which defines six levels of driving automation numbered 0 to 5. For AI agents, Feng, McDonald and Zhang (2025) proposed five levels defined by the user’s role: operator, collaborator, consultant, approver and observer, and argued that autonomy should be a design decision kept separate from capability. Practical levers include which actions need approval, spending and time budgets, reversibility and monitoring. Such scales help governance and procurement conversations, but levels blur in practice, because one agent may act freely on some actions and seek approval for others.
Related
Part of AI Terminology Index, section: Agents and Tools.