Autonomy levels for AI agents: deciding what runs without you
How much should an agent be allowed to do without asking? One answer for every agent and every action is always wrong. Levels let you give a little room first, and more when it has been earned.
Why one setting is not enough
Asking every time makes an agent useless, and people stop reading what they approve. Never asking works until the day the agent reads an instruction hidden in an email. The workable answer depends on two things: which agent it is, and what the action does.
Two dials: the agent and the action
The first dial is the agent's level: how far you trust this agent in general. A new agent, or one that reads a lot of outside content, starts low. The second dial is the type of action. Reading a calendar and deleting a mailbox are different risks whoever asks. A level decides which types an agent may do alone. No level should unlock everything.
- Read only: the agent can look and report. Nothing changes.
- Read and write: it can change your own things. Anything that leaves, or cannot be undone, waits.
- Trusted: it can also repeat what you have already approved, such as sending to a person you approved before.
What should never become automatic
Some actions deserve a person every time, at every level: sending to someone new, sharing outside, deleting, changing settings, setting up forwarding, moving money, and bulk actions. The cost of asking is a few seconds. The cost of not asking can be permanent.
Raising and lowering a level
Start every agent one level lower than you think it needs, and watch a week of its requests. If you approved the same thing every time, turn that one thing into a rule instead of raising the whole level: a rule is narrower. Lower the level, or stop the agent, the day it surprises you. Autonomy should be easy to take back.
Per agent, per workflow, per action
People often ask for autonomy per workflow. In practice a workflow is an agent plus the tools it uses, so the same result comes from three settings: a separate agent identity for each job, a level for that agent, and rules for the specific things it does often. Giving each job its own agent also means you can stop one without stopping the rest.
How Operon OS Agenaut does it
Each agent you connect has a level. Observer reads only. Assistant, the default, reads and writes, and anything else waits for you. Trusted is like Assistant and may send again to people you approved before; a new person still waits. Whatever the level, these always wait: sharing outside your account, sending to someone new, moving money, deleting, changing settings, new forwarding or filter rules, and anything done to more than 10 items at once.
A rule lets one agent do one thing without asking, with one person or with anyone. Agenaut offers a rule after you approve the same thing twice. Sending, paying and deleting can never be made automatic. Levels are set per agent, not per workflow.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- Human approval by action type: read, write, share, send, delete
- Human in the loop for AI agents: approve risky actions before they run
- AI agent governance: a practical checklist
Terms used here are explained in the glossary.