What is a semi-autonomous agent?

A semi-autonomous agent is an AI agent that completes some steps of a task on its own and hands others to a person, usually the steps that are risky, unusual or hard to undo. It plans and acts, but within a defined share of the work; the rest is drafted for approval, escalated or left alone.

Why does the "semi" matter?

Almost every agent in a regulated business is semi-autonomous, whether or not anyone calls it that. The useful question is which part is autonomous. The label on its own says little: an agent that sends routine reminders unaided and one that approves loans unaided but asks before sending reminders are both semi-autonomous.

So the term is a starting point, not a control. A supervisor will ask which actions the agent takes alone, on what evidence, and who answers for the rest. A label does not answer that. A map of actions does.

How does a semi-autonomous agent work in practice?

The agent receives a task, such as resolving a billing complaint. It reads the account, the complaint and the policy, then decides the steps. For each step, something outside the agent decides whether it may proceed:

• steps it may take alone, such as logging the complaint or sending an acknowledgement;
• steps it drafts for a person, such as a goodwill credit above a set amount;
• steps it must not take at all, such as closing the account.

The split should follow the cost of being wrong and the chance of reversal, not the agent's confidence. An agent that is sure of itself is still wrong sometimes, and the boundary exists for those times.

What does that look like in an example?

A building society's mortgage servicing agent handles requests to change the monthly payment date. It checks the account, confirms the new date is allowed under the product terms, and updates it on its own. If the change would leave a gap that looks like a missed payment, it drafts the change with a note and holds it for a servicing officer. If the customer mentions they are struggling to pay, it stops and routes the case to the team that handles financial difficulty.

One request, three possible outcomes, decided step by step. That is semi-autonomy working as intended.

How is a semi-autonomous agent different from an autonomous agent?

An autonomous agent completes a task end to end without a person approving any step. A semi-autonomous agent has at least one point where a person must act. The difference is not intelligence; the same model can sit behind either. It is how much of the work is allowed to complete unaided, and who decided that. The arguments for and against going all the way are in should AI agents be fully autonomous?

Which related terms should you know?

• Supervised autonomy: the agent acts while a person watches and can step in.
• Bounded autonomy: the agent acts alone inside limits enforced at the moment of action.
• Depth: how far an agent may act per capability, from reading to moving money.

How large the autonomous share should be for each task is the subject of how autonomous should an AI agent be?, and the principle behind the answer, that autonomy is a question about reversal, is set out in regulated autonomy.