Choosing the Right Level of Autonomy
A practical framework for deciding how much an agent may do on its own, from suggest-only to fully autonomous, based on risk and reversibility.
Every agent you deploy answers a question, whether or not you asked it: how much is this thing allowed to do without a person? Most teams answer it by accident. A prototype that sent emails directly "just for testing" goes live that way. A coding agent gets switched to skip-all-permissions on a Friday afternoon and never switched back. Or the opposite: an agent that's been right a thousand times in a row still waits for someone to click "Approve" on every step, and the team quietly stops using it.
Autonomy should be a decision, made per action, based on what happens when the agent is wrong. This lesson gives you a five-level ladder from suggest-only to fully autonomous, a set of questions for placing each action on it, hard ceilings that no amount of accuracy should lift, and a way to move actions up and down the ladder based on evidence rather than gut feeling.
The autonomy ladder
Like the numbered levels used for self-driving cars, a shared vocabulary helps. Here are five levels, defined by who executes the action and when a human looks at it.
| Level | Name | Agent does | Human does | Example |
|---|---|---|---|---|
| 1 | Suggest | Recommends what to do | Decides and does everything | "Here are three ways to reply to this complaint" |
| 2 | Draft | Prepares the action, ready to go | Reviews, edits, and executes it | A drafted email sitting in your drafts folder; a pull request you merge |
| 3 | Act with approval | Executes after a human approves | Approves each action or plan before it runs | "Send this email to Dana? Approve / Deny" |
| 4 | Act and report | Executes on its own within bounds | Reviews afterward, with undo where possible | Auto-labels and archives newsletters; posts a daily summary of what it did |
| 5 | Autonomous | Runs unattended within hard limits | Watches metrics, samples outputs, handles alerts | A nightly pipeline that triages new tickets and routes them |
The levels differ in more than convenience. At levels 1 and 2, the agent's mistakes are filtered through a human before they touch the world. At level 3, a human sees each action but under time pressure and with growing fatigue. At levels 4 and 5, mistakes reach the world first and get caught later, if at all. So the jump from 3 to 4 is the big one: it's where you stop preventing errors and start detecting and recovering from them.
You've already seen the ladder in tools you use. Claude Code's permission modes map onto it closely: Plan mode is level 1, Manual is level 3, Accept edits is level 4 for file edits, and Auto mode moves further toward 5 with a classifier checking each action. Permissions, approvals, and tool safety covers configuring them.