Autonomy is what an agent may decide. Authority is what it may change.
Why I keep an AI agent's freedom to think separate from its permission to act, and what that looks like in a real workflow.
Most agent setups hand the model a tool and, with it, the power to act. Deciding and doing arrive as one package. That's convenient, right up until the agent decides something you didn't expect.
Splitting the two
I split them. The agent can think as freely as it likes: read, compare, draft, propose. But the authority to change anything (write a record, send a message, move money) lives outside the agent, in rules a person wrote and can read.
When the agent wants to act, it asks. The rules answer with one of four decisions:
- Allowed: nothing risky is changed, so it goes ahead.
- Allowed within limits: it goes ahead only inside agreed bounds.
- Needs approval: it waits for a named person.
- Blocked: it was never given that permission, and the attempt is recorded.
The agent doesn't get to argue its way past that answer.
Why it matters for the people paying
If you're paying for finished work, you want to know the workflow can't quietly do something you never agreed to. Writing the limits down, and enforcing them outside the AI, is how that promise holds up.
Drafted with AI from my working notes, then reviewed and approved by me before publishing.