Let me start by saying, some people might immediately think this is nothing but clickbait. So I want to first establish what this article entails and the goal of writing it.
- Analyze the etymology of the word “agent” to see if it aligns with today’s agent marketplace.
- Identify various types of definitions, and categorize them for comparison.
- Compare those definitions for similarities in word use vs. buzzwords.
- Offer an aspirational, future-looking definition of an AI agent.
The Current Marketplace of AI Agents
Walk any trade floor this year, and you will hear the same word a thousand times. Agent. We have agentic platforms, agentic workflows, and agentic enterprises. The word has become a category, a valuation, a promise. And underneath all of it sits an awkward fact the industry would rather skip past: we have never agreed on what an agent is. Ask ten companies for a definition, and you will get ten answers, each one shaped around whatever they are selling.
This is the baseline of the problem. Agents are being positioned as marketing tools and hype vehicles, rather than as a standardized concept. A macro became an agent. A chatbot became an agent. A piece of code that performs an automated task is now an agent. Every major vendor has unlocked the magic of agents, prompting all to see on their home pages. And that was over a year ago.
When an agent expands to cover anything, we have nothing to compare it to. We have agent washing. We have hype. We have the largest marketing campaign in human history to justify the claim that we have code entities that can make your work easier. We’ve always had those opportunities, but now you can give those opportunities names and identities and hire them as employees.
My toaster is now named Toastimus Prime with a cryptographically signed ID.
The Origins of “Agent”
It helps to go back to where the word began. Agent comes from the Latin agere, to act, to drive, to set in motion. But the word never meant simply one who acts. In law and in commerce, where it matured, an agent is one who acts on behalf of someone else.
A principal sets the goal.
The agent carries it out under the granted authority.
And when the agent acts, the responsibility runs back up the chain of command to the principal.
Defined as:
Agent /ˈāj(ə)nt/ (noun)
An agent is a person, entity, or substance authorized or capable of acting on behalf of another to produce a specific effect.
In business, legal, and medicine, we extend this to areas such as:
Real Estate Agent: Helps clients buy, sell, or rent properties.
- Travel Agent: Arranges transportation and lodging for clients.
- Talent/Sports Agent: Represents actors, athletes, or artists to secure employment and contracts.
- Secret Agent: An undercover spy working for an intelligence organization.
- Chemical/Biological Agent: A substance or organism that produces a specific physical or chemical change (e.g., a cleansing agent)
The word was never a measure of capability. It described a relationship and a chain of accountability inside that relationship.
Yet within AI, the industry began to shift that accountability.
- AI/Software Agent: A computer program designed to act autonomously or automate specific tasks (like gathering data).
Even if we want to argue that this is acting on behalf of another, there is a key contradiction within this definition related to how the agent “acts” towards a task. More on this below.
Current “AI Agent” Definitions
The oldest definition of system agents is an academic version.
Michael Wooldridge and Nicholas Jennings (1995) define an intelligent agent as “a computer system situated in some environment that is capable of flexible, autonomous action to meet its design objectives.”
Russell and Norvig defined a rational agent as "anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators."
Unfortunately, by their standards, a smart thermostat is considered an agent.
The newer definitions are more engineering and marketing versions, generated due to the popularity of LLMs.
IBM defines an AI Agent as:
- An artificial intelligence (AI) agent is a system that autonomously performs tasks by designing workflows with available tools.
Anthropic defines an AI Agent as:
- An intelligent system that autonomously directs its own processes and tool usage to accomplish complex tasks.
LangChain defines an AI Agent as:
- A system that uses an LLM to decide the control flow of an application.
Google defines an AI Agent as:
- An intelligent software system capable of autonomously understanding goals, planning multi-step actions, and executing tasks on your behalf.
The Massive Problem with These Definitions
There’s one word that recurs across these definitions and points to the core issue.
The word is autonomous.
We are constantly told that an agent is an autonomous system. I want to argue that this single word does more to corrupt our definition of an agent than any other.
There are two problems with it, and they compound. The first is simple and empirical. No system in production today is autonomous.
Autonomy means a system that originates its own ends, that sets its own goals and governs itself. Every system shipping today runs on goals a human handed it. So if we insist that an agent must be autonomous, we have written a definition that nothing today can satisfy, and we have proven, in our own wording, that no agents exist.
If an agent is autonomous, and no system is autonomous, then we have no agents.
The second problem is deeper and lives within the words and definitions themselves. Autonomy comes from the Greek autos, self, and nomos, law. To be autonomous is to be a law unto yourself.
In contrast, an agent acts under the law of another. Our earlier definition established that. An agent pursues a principal’s goals under a principal’s authority.
There is a Greek word for that condition, too: heteronomy, from heteros, "other." An agent is heteronomous by definition. Serving another’s purpose is the whole of the job. Which means the phrase "autonomous agent" asks for a single thing: to be self-governed and other-governed at the same time.
The more autonomous a system truly became, the less it would serve as anyone’s agent, because it would have started answering to itself.
In all cases, the definitions refer to automation. Autonomy speaks to authorship.
Today’s AI Flavor of the Month
There is one more point of confusion to clear up. We assume an agent must be made of a large language model. That assumption owes more to familiarity than to fact. The language model is the part of modern AI that a person can see, touch, and talk to, so it became the face of the whole agent campaign. Basically, no one knows any different. An LLM vs. an Analytical model is the same in about 90% of the market’s minds.
The planners, control systems, and learning algorithms that run logistics, markets, and games pursue their objectives far more relentlessly than any chatbot. The language model is the interface, the part that talks. It is rarely the part that drives. A definition welded to it inherits its expiration date. Today, we say "agent" and picture a language model only because that's what we happen to have.
But what happens when LLMs go away?
A Stronger Definition for AI Agents
If we wanted to define the agent properly, and to build the thing the word has promised since Rome, we already know the shape it would take. Strip away the marketing, and the requirements fall out cleanly. A real definition includes the clauses that today's versions omit entirely.
Chris Hood's AI Agent Definition
- An AI agent is a persistent, identifiable software entity that holds delegated authority to pursue goals on behalf of a principal, acts and transacts with other agents and systems under that authority, and whose actions remain attributable to that principal.
Look closer at that definition. Every clause is a tether. Identity binds the system to a name. Delegation binds it to a grantor. Accountability binds it to a principal. And notice the word that has gone missing. Autonomy. A legitimate agent is defined by what holds it, rather than by how freely it runs. The dream the market keeps selling, a system finally free of all restraint, would fail every clause at once. Which turns the usual pitch on its head. We should want agents that are exquisitely governed rather than gloriously free. In honest terms, autonomy is the failure state. Heteronomy, acting faithfully for another, is the entire definition.
Here is the part that should make us laugh, or wince. None of this is new. Identity, delegated authority, accountability, and bounded discretion. That is the principal-agent relationship, and lawyers have understood it for centuries. We never needed to invent a definition of an agent. We needed to remember the one already sealed inside the word the day we borrowed it.
That’s the clearest sign of hype than anything else.
My definition is aspirational, and I will own that. Almost nothing on the market meets it. But that is the point of a definition worth having. It should describe the thing we are trying to build, and hold the line against everything we are tempted to call by its name too soon.
