What is an AI agent?

Two things are true about this word at once. It has a thirty-one-year-old definition that most of the field still accepts, and it has no definition at all in the law that governs what these systems are allowed to do. Anybody choosing between products, writing a policy or deciding who signs off a deployment is working in the gap between those two facts.

Four properties, settled in 1995#

Michael Wooldridge and Nicholas Jennings opened their 1995 survey in The Knowledge Engineering Review by reporting a complaint that has aged unnervingly well. Carl Hewitt, they write, had remarked that the question "what is an agent?" embarrasses the agent-based computing community in the way "what is intelligence?" embarrasses mainstream AI, because the term "defies attempts to produce a single universally accepted definition". They added a warning: "unless the issue is discussed, 'agent' might become a 'noise' term, subject to both abuse and misuse, to the potential confusion of the research community."

Their answer was a property list, which they called the weak notion of agency. In their own words, an agent has autonomy, operating "without the direct intervention of humans or others" with "some kind of control over their actions and internal state"; social ability, interacting with other agents and possibly humans "via some kind of agent-communication language"; reactivity, perceiving an environment and responding "in a timely fashion to changes that occur in it"; and pro-activeness, meaning agents "do not simply act in response to their environment, they are able to exhibit goal-directed behaviour by taking the initiative".

A stronger notion, which they flag as contentious, adds the vocabulary of minds: knowledge, belief, intention, obligation. Most of what is sold today claims the weak notion and is described in the language of the strong one.

The line that does the operational work#

For anyone buying or building, the useful distinction was published by Anthropic on 19 December 2024, and it turns on a single question. "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." Both sit under the heading agentic systems, and the post opens by conceding that "'Agent' can be defined in several ways".

Apply that line to a shortlist and most of it moves. A support tool that classifies a ticket, looks up an account and drafts a reply has a human-written sequence with a model at each junction. It is a workflow wearing the word. This is not a criticism of workflows, which are usually the better choice: they are predictable, testable and cheap to reason about. It is a criticism of the label, and the label is what a buyer is being quoted a price against.

A search of the EU AI Act that returns zero#

Regulation (EU) 2024/1689 is the most detailed body of binding AI law anywhere. The full Official Journal text runs to 588,616 characters, from the title through Annex XIII, and it was read and searched on 22 September 2026. The word "agent" appears in it zero times. So does "agentic". As a control on whether the text had loaded completely, "AI system" returns 1,107 matches. The 21 hits for "agency" were inspected one by one and every one of them is either an institutional body, such as the European Union Aviation Safety Agency, or the ethics-guidelines phrase "human agency and oversight".

What the Act does define, at Article 3(1), is an AI system: "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments". That wording closely tracks the definition the OECD adopted on 8 November 2023, which likewise has no agents in it. The OECD's term for whoever acts is "AI actors", and its definition covers "organisations and individuals that deploy or operate AI". The acting party, in both instruments, is a person.

None of this means agents escape regulation, and a page that implied so would be wrong. The obligations attach to the system and to the provider or deployer placing it on the market, whatever the marketing calls it. What it means is narrower and more useful: no legal consequence anywhere in that Regulation turns on whether a thing is an agent. A procurement argument about whether a product qualifies is a commercial argument, not a compliance one.

The only body that has defined the term cannot compel anybody#

This page was drafted expecting to report that no standards body had defined the term. That was wrong, and checking it produced the better finding. ISO/IEC 22989:2022 defines "AI agent" at clause 3.1.1, which is the very first defined term in the standard, ahead of "AI system" at 3.1.4. Its wording: "automated entity that senses and responds to its environment and takes actions to achieve its goals", where "automated" is separately fixed at 3.1.7 as functioning "without human intervention" under specified conditions.

Two things about that definition are worth carrying. It predates the current wave, so it was written about robotics and control systems and inherits their assumptions. And the same standard defines an AI system as one that generates outputs "for a given set of human-defined objectives", which is the pre-2023 formula both the OECD and the EU have since moved away from. So the one document that defines the agent is anchored to a description of AI that its neighbours have abandoned. Its bar for "autonomy" is higher again, requiring a system "capable of modifying its intended domain of use or goal without external intervention", which almost nothing on sale today meets.

The gap is live and the standards bodies know it. NIST's Computer Security Resource Center glossary has no entry for "AI agent" at all, defining only "agent" in an adversarial machine-learning security context as "software programs that can interact with their environment, receive information, and undertake self-directed actions in service of a larger, externally-specified goal". NIST opened an AI Agent Standards Initiative in February 2026 to convene industry towards standards that do not yet exist. The UK Competition and Markets Authority, writing on agentic AI in March 2026, simply says "Definitions vary" and offers a cluster of capabilities instead: "a degree of autonomy from continuous human supervision; goal-orientation, where systems pursue outcomes rather than isolated tasks; multi-step reasoning across complex decisions; and potentially the ability to act across systems, platforms and data sources".

Why the word is worth pinning down at all#

The estate's interest here is not taxonomy. It is that the definition decides where a person has to stand.

Under a workflow, the sequence was written by somebody who can be named, and every branch in it is a decision a human took in advance and can be asked about. Under an agent, on Anthropic's own line, the system chooses its own route, so the human decision moves from the path to the goal and the guardrails. Those are different jobs requiring different competence, and an organisation that has bought one while staffing for the other has a gap it cannot see. That gap is the subject of who manages AI agents and of the moral crumple zone, where accountability settles on whoever was nearest the failure rather than on whoever designed the conditions for it.

The vaguer the word, the easier that drift becomes. Wooldridge and Jennings saw the mechanism in 1995 and called it a noise term. Thirty-one years later the noise has a procurement budget attached.

Three questions that turn the word back into a description#

Who picks the next step? If a person wrote the sequence, it is a workflow and the testing burden is ordinary software testing. If the model picks, the sequence is different every run and cannot be tested the same way.

What can it do without asking? Autonomy is the property with the most range in it. Reading is not writing, drafting is not sending, and recommending is not spending. Every serious definition on this page, from 1995 to ISO to the CMA, treats autonomy as a scale, so a policy that treats it as a switch is answering a question nobody asked.

Aim your definition at the answer to those first two and the third follows. Who is answerable when it acts? There is a named provider and a named deployer under the Act whatever the product is called, so the question always has an answer, and the only variable is whether that person knew they were the answer. Related reading: whether to let an agent act on your behalf, and what the rogue agent incidents mean.

Key sources

• Wooldridge, M. and Jennings, N. R. (1995). Intelligent agents: theory and practice. The Knowledge Engineering Review, 10(2), 115-152.
• ISO/IEC JTC 1/SC 42 (2022). ISO/IEC 22989:2022, Artificial intelligence concepts and terminology, clause 3.1.1.
• European Parliament and Council (2024). Regulation (EU) 2024/1689, Article 3: Definitions. Official Journal of the European Union.
• OECD (2023). Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449, definition of an AI system as revised 8 November 2023.
• Anthropic (2024). Building effective agents. Anthropic Engineering, 19 December 2024.

Related SuperSkills research#

On the human side of a deployment, who manages AI agents and how humans and agents divide work. On delegation and its limits, letting an agent act on your behalf and AI agents and human judgement. On what happens when one goes wrong, the rogue agent incidents and the moral crumple zone. On uncertainty, how an agent should communicate it.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Every definition on this page belongs to somebody else and is credited: the four properties to Wooldridge and Jennings, the workflow line to Anthropic, clause 3.1.1 to ISO/IEC, Article 3(1) to the European legislature. Nothing here is a SuperSkills coinage. All five sources were read at their primary locations on 22 September 2026, and the EU AI Act count was produced by searching the full Official Journal text rather than a summary of it. The page originally asserted that no standards body defines the term; ISO/IEC 22989 falsified that and the claim was narrowed to binding instruments. Last reviewed: 22 September 2026.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.