Governance
How much control you hand to an agent is a design choice, not a switch. Five levels, from operator to observer, told as a walk with a small robot.
Nov 7, 2026 · 4 min read
Teams often talk about agents as if there were two modes: a chatbot that answers, and an autonomous system that does everything. Real projects live in between, and the useful question is not whether to use an agent but how much it should decide on its own, for which task.
A 2025 paper, Levels of Autonomy for AI Agents, argues that an agent's level of autonomy is a deliberate design decision, separate from how capable the agent is. It defines five levels by the role the person takes: operator, collaborator, consultant, approver and observer. The walk below gives each level a plain-language reading of our own.
You make every decision and the agent does what it is told. A good way to learn what it can and cannot do.
You and the agent plan and do the work together, and either of you can take the lead on a step.
The agent takes the lead and comes to you when it needs your expertise or a preference only you know.
The agent prepares the work and stops at the moments that matter, such as spending money or writing to a customer.
The agent acts on its own inside limits you set, and you monitor and can step in at any time.
Tell us about one process. We will show how an agent could handle it, and what to measure to know it works.