Why the phrase contradicts itself, and why that contradiction keeps the field from defining what an AI agent really is.

"Autonomous agent" may be the most-hyped phrase in technology this year (second only to "agent identity"), and also one of the most incoherent, the kind of term a marketing team ships before checking whether it means anything. The two words pull against each other. An agent acts for someone. An autonomous thing answers to itself. Put them together, and you get something that cannot exist, and a fresh layer of confusion on an industry that was already confused.

What an "AI Agent" Actually Is

Let's start with "agent" as a noun. The word comes from the Latin agere, to do, to drive, to act. An agent acts for a principal. The real estate agent sells your house. The literary agent shops your manuscript. The diplomat acts for a state. Law built an entire doctrine on that relationship: agency is the arrangement in which one party acts on behalf of another, under that other's authority, toward that other's ends.

We give an AI agent a goal. It pursues the goal. We own the end; it works the means. Remove the marketing, and that relationship is the whole definition of an agent.

An agent is a person, entity, or substance authorized to act on behalf of another, or capable of producing a specific effect. In this case, we declare that the entity is an AI, acting on your behalf.

"Autonomy" and the Meaning of Self-Law

"Autonomy" combines the Greek autos, "self," with nomos, "law." It translates literally into self-law. An autonomous thing carries its own legislation. Kant set this at the center of his moral philosophy and paired it with a precise opposite, heteronomy: being governed from outside, taking your law from another hand. It drew the line between a will and a mechanism, between a thing that authors its own ends and a thing that receives them.

Autonomy is the right to, or condition of, self-government: independence, and the capacity to make one's own choices free of external control.

Heteronomy is the opposite condition, in which a thing is governed, directed, or moved by an outside power rather than by itself.

These two definitions lay the foundation for the problem with the term "autonomous agent." An agent receives its directions from an outside power. An autonomous entity authors its own.

Why "Autonomous Agent" Is an Oxymoron

Look at how the industry uses the term "autonomous agent," and the contradiction shows up on its own.

A widely cited definition from a recent economics paper on AI transactions calls agents autonomous software systems that perceive, reason, and act to achieve goals "on behalf of human principals." The same sentence says the system governs itself and that it works for a principal. A system cannot take orders from you and write its own orders at the same time.

An older, equally cited definition from Franklin and Graesser goes further. It has the agent acting in pursuit of "its own agenda." An agenda of its own is the one thing the word agent was built to rule out, because an agent exists to carry out someone else's agenda. The definition hands the agent the one thing that would stop it from being an agent.

An agent takes its law from the principal, which is the definition of heteronomy. Calling that same agent autonomous asks it to take its law from itself in the same breath. You end up describing a self-governing servant, a thing that answers only to itself while existing only to answer to you. That is an oxymoron in the strict sense, and the strain shows up every time someone tries to pin the term down.

The funnier reality, and one you may have already reached yourself, is that the honest correction makes things no better. "Heteronomous agent" is the accurate phrase, but it only trades one problem for another. It slides the term from oxymoron to pleonasm, from a contradiction to a redundancy, because an agent is already governed from outside. The accurate version says the same thing twice. The hyped version says two things that cannot both be true. There is no version of the phrase that lands cleanly, which should tell us something.

Why the Industry Cannot Agree on a Definition for "AI Agent"

The field admits, in its own surveys, that the term is nebulous, with definitions scattered across cybernetics, reinforcement learning, software engineering, law, and philosophy that rarely line up. We tend to wave this off as a young field still finding its feet. The deeper cause is the oxymoron.

The EU AI Act declines to define an agent at all and regulates systems that operate at "varying levels of autonomy." Researchers propose autonomy ladders, six rungs running from none to full. One careful paper concedes that it uses "agent" and "autonomy" almost interchangeably, since autonomy is the only knob its experiments turn. None of those is a definition. Each is a workaround. When your central term conceals a contradiction, you stop defining it and start grading it, because a clean definition would drag the contradiction into the open.

The oxymoron does more than offend a logician. It blocks the field from saying plainly what an agent is, because the instant you try, the borrowed adjective pulls the sentence apart.

We also have to recognize that most definitions of an AI agent today rest either on an individual's narrow understanding of AI or on an organization's marketing perspective. If it sells more widgets, then everything becomes autonomous, even though no system is autonomous.

Two Common Defenses, and Why Both Fail

Two defenses usually arrive here.

The first: an agent has discretion, so it is partly autonomous. Yet discretion lives inside a mandate. A lawyer exercises judgment on your instructions that are never spelled out, and she stays your agent only as long as she serves your ends. Let her pursue her own, and she has stopped being your agent and become a liability. Latitude over the means leaves ownership of the ends exactly where it sat. And latitude is granted, which means it can be revoked, which makes it the reverse of self-law rather than a small dose of it.

The second: computer science has already redefined "agent" to mean a system that senses and acts in an environment, with the principal removed. True, and revealing, because draining the principal out of the words is what lets autonomous slip in looking innocent. The principal left the sentence. It stayed in the work. We still build these systems for our goals, still call them ours, still expect them to do what we meant them to. The vocabulary forgot the relationship. The relationship held.

The bottom line is that the industry has bastardized two words, Agent and Autonomy, and then made things worse by fusing them into a single term that makes no sense to anyone who understands how these systems actually work.

Authority Laundering and the Oxymoron

Call a system a tool, and its maker owns the outcome. Call it autonomous, and the outcome floats free, with no one on the hook. I have called that maneuver "authority laundering": running a human decision through enough machine-sounding vocabulary that the human fingerprints rinse off. Someone chose the goal. Someone set the constraints. Someone shipped it. Autonomy talk takes that chain of choices and dissolves it, so that when the system acts, the action appears to come from nowhere, with no person left to name.

Agent washing is the same move at the level of the product label. Slap "autonomous" on the box and a piece of software starts to look like a free actor rather than a delegated one, which is exactly the impression a vendor wants when the question of liability comes up. An undefined "autonomous agent" turns out to be convenient precisely because an undefined thing is hard to hold responsible. The hype and the vagueness are the same features seen from two angles. One sells the product. The other shields it.

A Cleaner Definition of an AI Agent

An agent is a system that acts on behalf of a principal, toward a goal the principal sets, with latitude over the means. That definition is clean, old, testable, and it describes exactly what every lab is building. It tells you where accountability lives, because it names a principal. It scales from a thermostat to a frontier model without buckling. And it asks for nothing self-contradictory, because it never pretends the thing governs itself.

Between human-in-the-loop carve-outs and various "bounded autonomy" justifications, the industry has chased the hype around something that does not exist. No system is autonomous, because no system understands itself. The belief in these terms survives because it sells more agents, which is a hard thing to do when we have no shared definition of an agent in the first place.

So the next time someone hands you a definition of an AI agent, ask whether they can define it without the word "autonomous."