Glossary

Autonomous agent

Autonomy is a range, and the interesting question is where the boundary sits.

Definition

An autonomous agent is a system that pursues a goal by deciding its own sequence of actions, rather than executing a procedure authored in advance. Autonomy is a matter of degree, bounded in practice by the tools the agent can use, the limits placed on its run, and the approval points where a human decision is required before it can proceed.

Research origin: Multi-agent systems, 1990s
Property described: Decision latitude
Bounded by: Tools, limits, approval gates

The term is older than the current wave

Autonomous agents were a research field long before language models. Wooldridge and Jennings' 1995 survey of intelligent agents in the Knowledge Engineering Review set out the properties still cited today: autonomy, reactivity, proactiveness and social ability. Robotics, simulation and distributed AI used the vocabulary for decades.

What changed after 2022 is not the definition but the substrate. Deciding what to do next used to require hand-built planners and domain models; a language model with tool access does it directly from a natural-language goal. The old definition survived the change intact, which is why the research usage and the product usage still line up.

Degrees of autonomy in practice

Most product arguments about whether something is truly autonomous are arguments about which row it occupies.

Level	What the system decides	What a human decides
Scripted	Nothing. It follows the authored path	Every branch, in advance
Tool-selecting	Which tool to call next, within a fixed procedure	The procedure and its order
Goal-pursuing	Its own steps, tools and stopping point	The goal, the criteria, and acceptance of the result
Self-directed	Which goals to pursue at all	Policy and scope, if anything

Commonly confused with

Term	What it describes	The difference
Headless agent	The interface: no chat window, triggered programmatically	An orthogonal property. Agents can be headless and scripted, or interactive and highly autonomous.
Automation	Executing a defined procedure without human effort	The path is authored. Autonomy is about choosing the path.
AGI	General intelligence across arbitrary domains	A capability claim about breadth. Autonomy is a design property about latitude within a task.
Self-improving system	A system that modifies its own capabilities	A different axis entirely, and not implied by autonomy.

Where the boundary is usually drawn

Production systems almost never sit at the self-directed end, and the reason is accountability rather than capability. An agent that picks its own goals has no natural point at which a person is answerable for the outcome. So the common design is goal-pursuing autonomy inside a task, with the goal and the acceptance criteria set by a human and the result accepted by one.

Polaris sits exactly there. An AI worker chooses its own approach to a task, runs its own searches and decides when the criteria are met, and it cannot mark the task done. A person closes it and rates the work.

Headless agent

Nobody is typing at it, so its output is actions and artefacts rather than replies.

Human-in-the-loop

The system cannot complete the loop without a person, by design.

Agent orchestration

Deciding which agent does what, in what order, and what happens when one fails.

AI worker

The difference between an agent you prompt and an agent you assign work to.

Acceptance criteria

Written before the work, checkable after it, and binary either way.

Assigning a task to an AI worker

There is no prompt box. The task is the prompt, and the checklist is the contract.

Questions people ask

+Is an autonomous agent the same as a headless agent?

No. Autonomy describes how much latitude the agent has over its own actions; headless describes whether it has a user interface. The two vary independently, and treating them as synonyms is the most frequent mistake made with both terms.

+How autonomous should an agent be?

As autonomous as the reversibility of its actions allows. Drafting a document, running research and producing a file are cheap to undo and can be left to the agent. Sending external messages, moving money and deleting records are not, and belong behind an approval step regardless of how capable the agent is.

+Who is responsible for what an autonomous agent does?

The organisation that deployed it, in every legal framework currently in force. That is the practical reason approval gates and audit records exist: responsibility does not transfer to the software, so the humans holding it need a way to see and control what was done.