An autonomous agent is an AI agent that pursues a goal and takes actions in other systems without a person approving each step. It decides what to do next, does it, checks the result and carries on. Autonomy is not all or nothing: the useful question is never whether an agent is autonomous, but which actions it may take alone, within what limits.
What makes an agent autonomous?
Three things. It chooses its own steps rather than following a fixed script. It acts through tools that change records, send messages or move money. And it does so without waiting for a person between steps. Remove the third and you have an assistant; remove the second and you have a chatbot.
In practice, most agents in regulated businesses are autonomous for some actions and not others. An agent might update records on its own all day and still hold every payment for a person.
Why does autonomy raise the stakes?
Because errors are no longer filtered. When a person approves each step, a wrong suggestion is caught before it does harm. When the agent acts alone, a wrong step happens, and the next step builds on it. Autonomy multiplies the speed at which an agent works and the speed at which a mistake spreads.
That does not argue against autonomy. It argues for granting it where mistakes are cheap and reversible, and earning it with evidence where they are not. How autonomous an AI agent should be works through that trade-off.
How is autonomy granted in practice?
Capability by capability. For each action an agent could take, the business decides the most autonomy it would ever allow, tests the agent on its own real cases, and grants autonomy up to that limit only where the evidence supports it. Autonomy is checked at the moment of action, every time, and it expires, so it has to be re-earned.
For example, a mid-sized insurer's claims agent acts alone to request missing documents, update claim status and schedule assessor visits. It drafts settlement offers but cannot send them. It cannot decline a claim at all; declines go to a handler. The agent is autonomous in a defined space, and the edges of that space are written down.
How is an autonomous agent different from a semi-autonomous agent?
A semi-autonomous agent acts on its own for some steps and needs a person for others, usually at fixed points. In practice this is how most autonomous agents in regulated work operate, so the line is blurry. The useful difference is whether the agent's independence is set per action, on evidence, or granted as a blanket label. A blanket "autonomous" label invites over-trust; per-action autonomy can be shown and defended.
Should AI agents ever be fully autonomous?
For narrow, low-risk, reversible work, an agent can act alone for long stretches. For actions that move money, affect people's rights or cannot be undone, keeping a person at the point of no return is usually right, whatever the agent's track record. The full case is in should AI agents be fully autonomous? and the principle behind it in regulated autonomy.
Where Pryme fits
Pryme grants autonomy as depth, per capability, per customer, on evidence from the customer's own cases. Depth is enforced at the moment of action, certificates lapse after 90 days, and some actions are withheld for every agent. The verdict on every action is filed, so the edges of an agent's autonomy can be shown, not just described.
Questions people ask
Is an autonomous agent the same as an AI agent?
Every AI agent has some autonomy, because it chooses its own steps. "Autonomous agent" usually stresses that it acts without step-by-step approval.
Who is responsible when an autonomous agent makes a mistake?
The business that deployed it, through the named people who own the process and approved the agent's autonomy. Autonomy shifts who does the work, not who answers for it.