Autonomy Is a Dial, Not a Switch

People tend to ask whether an AI agent is "autonomous" as if it were a yes-or-no fact about the tool. It is closer to a dial you set. The same agent can ask permission for every move or run a whole task and report back — and where you turn the dial is your call, tuned to how much a mistake would cost.

The three common settings

• In the loop — the agent proposes, you approve each action before it happens. Most control, most of your time. Right for irreversible or high-stakes work.
• On the loop — the agent acts on its own, you watch and can step in. It opens the pull request; you review before it merges. The common setting for real work.
• Out of the loop — the agent runs end to end with no one watching in real time. Reserve it for low-stakes, easily reversible tasks.

Most disappointment, and most danger, comes from a mismatched setting: turning the dial to "out of the loop" on work that needed a human in it.

Who is accountable

Here is the part the word "autonomous" hides: the agent is never accountable. It does not own the outcome, feel the consequence, or get fired. When an agent acts, the responsibility sits entirely with the people who deployed it and set the dial. Treat its output like a capable contractor's draft — useful, and yours to check before it counts.

How to choose the setting

One question sets the dial: how bad is the worst case, and how hard is it to undo? Cheap and reversible — turn it up, let the agent run. Expensive or permanent — turn it down, approve each step. The skill is matching the dial to the stakes, task by task.

That matching has a home in the pillars: Permission Architecture is how you build the guardrails the dial relies on.