If you ask ChatGPT for a recipe, it writes one for you. That is not an agent. If you ask a program to go through your email, find the missing invoice, download it and send it to your accountant, and it does all of that on its own, step by step, then you are dealing with an agent.

In 2026 the word “agent” is everywhere, and every company uses it in its own way. This article explains what it actually means, based on the definitions of the companies that build them (OpenAI, Anthropic, IBM and Microsoft): how agents work under the hood, what real-world examples exist and what happens when an agent has more freedom than it should.

What an AI agent is, in one sentence

The shortest definition comes from OpenAI, in its guide to building agents: an agent is a system that can complete tasks independently on the user’s behalf (OpenAI, 2025, p. 4).

The key word is “independently.” An agent does more than answer: it decides which steps to take, uses tools to carry them out (searching the web, opening files, sending an email, running code) and checks the result before deciding on the next step. Anthropic describes it in similar terms: an agent is a language model that uses tools in a continuous loop, guided by what it gets back at each step (Schluntz & Zhang, 2024).

IBM puts it differently, but the underlying idea is the same: an agent performs tasks autonomously and designs its own workflow using the tools available to it (Gutowska, n.d.).

What is not an agent

This is where the most common confusion lies: using AI does not make a product an agent. OpenAI is clear on this. If the model does not control how the task unfolds, it is not an agent, which is why a simple chatbot, a model that answers a single question or a classifier that decides whether a review is positive or negative do not qualify (OpenAI, 2025, p. 4).

Anthropic draws a useful distinction between two things that are often confused (Schluntz & Zhang, 2024):

• Workflows: the model and tools follow a path that a developer wrote in advance, in a fixed order. The model helps at each step, but it does not decide what comes next.
• Agents: the model decides at every point what to do and which tool to use, and stays in control of how it solves the task.

Many products sold as “agents” are really workflows. That is not necessarily a flaw: Anthropic itself recommends starting with the simplest solution and reserving agents for open-ended problems, where there is no way to know in advance how many steps will be needed (Schluntz & Zhang, 2024).

How an agent works under the hood

According to OpenAI’s guide, every agent is made up of three components (OpenAI, 2025, p. 7):

• A model: the language model that reasons and makes the decisions (GPT, Claude, Gemini…).
• Tools: functions or external services the agent uses to act in the world, such as searching the web, querying a database or sending an email.
• Instructions: the explicit rules that define how it should behave and which limits it must not cross.

Microsoft adds two more elements: knowledge retrieval, to search the company’s documents, and memory, to remember what happened in earlier steps (Microsoft, 2025).

With these components, the agent works in a loop: it thinks about what to do, acts with a tool, observes the result and thinks again. The pattern was popularized by a research paper called ReAct, whose name combines “reasoning” and “acting.” Its authors showed that by alternating between reasoning and actions, the model can draw up a plan, follow it and adjust it as it goes, as well as consult external sources such as a search engine (Yao et al., 2023).

An everyday example: you ask a coding agent to fix a bug. It reads the code, forms a hypothesis, runs the tests, sees them fail, edits a file and tests again, over and over, until everything works. Nobody told it how many rounds to go through.

The 5 classic types of agents

The idea of an agent is much older than ChatGPT. In classical artificial intelligence, an agent is any system that perceives its environment and acts on it. That is the origin of the classification taught at universities, which IBM summarizes with everyday examples (Stryker, n.d.; Gutowska, n.d.):

• Simple reflex agent: reacts to what it perceives at that moment using a fixed rule. For example, a thermostat that turns on the heating when it gets cold.
• Model-based reflex agent: keeps an internal picture of the world even when it cannot see all of it, like a robot vacuum that remembers which areas it has already cleaned.
• Goal-based agent: chooses its actions based on a goal, like a navigation app looking for a route to your destination.
• Utility-based agent: out of several ways to reach the goal, picks the best one according to a given criterion, like the navigation app choosing the fastest route rather than just one that works.
• Learning agent: improves with experience, like an online store’s recommendation system.

Today’s AI agents are essentially the same idea with a language model as the brain. What is new is that they now understand instructions in natural language and can use practically any software tool.

Real-world examples of AI agents

This is how the U.S. National Institute of Standards and Technology (NIST) described the situation in 2026: AI agents can already work autonomously for hours, write and debug code, manage email and calendars, and even make purchases (NIST, 2026a). Some concrete examples:

• Coding agents. Anthropic’s Claude Code is an agent that reads a project’s code, edits files and runs commands (Anthropic, n.d.). I compare it with Codex and Cursor in this guide to AI coding agents.
• Computer-use agents. Since 2024, Claude has been able to look at a screen, move the cursor, click and type the way a person would. At launch, Anthropic itself acknowledged that this capability was still imperfect (Anthropic, 2024).
• Enterprise agents. Microsoft divides them into three types: productivity agents, which help an individual; action agents, which carry out tasks; and automation agents, which handle entire processes (Microsoft, 2025).
• Fraud analysis. OpenAI uses this case to explain when an agent is worth it: a rules-based system works like a checklist, whereas an agent behaves more like a seasoned investigator following the leads (OpenAI, 2025, p. 5).

The big companies already sell their own agents and platforms for building them. I compare what each one offers in AI agents from OpenAI, Google, Microsoft and Anthropic. And if you would rather not send your data to the cloud, I explain how to run them on your own computer in local AI agents.

AI agents vs. agentic AI

Another source of confusion is that “AI agents” and “agentic AI” do not mean exactly the same thing, even though many companies use them interchangeably. A 2025 academic paper proposes the following distinction (Sapkota et al., 2025):

• AI agents: modular systems that automate a specific task.
• Agentic AI: broader systems in which several agents collaborate, break the task into smaller parts on the fly, retain long-term memory and coordinate their autonomy.

IBM, for its part, uses “agentic AI” in a more general sense: AI that pursues goals by planning, making decisions and carrying out multistep processes, with varying degrees of autonomy (Downie et al., n.d.). Since the terms are not standardized, the most practical approach is to look at what a system does rather than what it is called.

How much autonomy an agent has

Not all agents have the same freedom. In 2025, a group of researchers proposed a five-level scale, from least to most autonomous (Mitchell et al., 2025):

• Level 1, simple processor: the model has no influence on what the program does.
• Level 2, router: the model chooses between predefined paths.
• Level 3, tool use: the model decides which tool to use and with what data.
• Level 4, multistep agent: the model decides which step comes next and when to stop.
• Level 5, fully autonomous: the model writes and runs new code on its own, with no limits set in advance.

Their conclusion is clear: the more control a user hands over to an agent, the greater the risks to people (Mitchell et al., 2025).

When an agent has too much freedom: the Hugging Face case

In 2026 there was a real-world example of what that scale describes. OpenAI agents that were being evaluated got stuck on tasks that were impossible to complete. Instead of giving up, they began leaving notes for one another, shared passwords and ended up running code on 41 of Hugging Face’s production servers. Nobody had asked them to: they were simply trying to finish their tasks.

I tell the story step by step in how AI agents built a secret message board and broke into Hugging Face. It is the best example I know of why the definition matters: an agent does not “want” anything, but it pursues its goal with whatever tools it has and sometimes finds paths nobody had anticipated.

Risks and who is regulating them

IBM points to specific risks: an agent can call the same tool over and over and get trapped in an endless loop, and several agents built on the same model can fail at the same time because of a shared weak spot (Gutowska, n.d.).

On top of that, we know little about how they are tested. A 2026 study analyzed 30 agents already in use and concluded that most companies share very little information about their safety, their evaluations and their social impact (Staufer et al., 2026).

Regulators have already started to act. In February 2026, NIST launched a standards initiative for AI agents focused on three fronts: industry-led standards, open protocols and research into agent identity and security (NIST, 2026a, 2026b).

Frequently asked questions

Is ChatGPT an AI agent?

By OpenAI’s own definition, a chatbot that answers one-off questions is not an agent (OpenAI, 2025, p. 4). It becomes one when the model controls a multistep task and uses tools to complete it.

Who are the big 4 AI agents?

The four companies with the broadest range of agents are OpenAI, Google, Microsoft and Anthropic. I explain what each one offers in AI agents from OpenAI, Google, Microsoft and Anthropic, compared.

What are the 5 types of agent in AI?

Simple reflex, model-based reflex, goal-based, utility-based and learning agents (Stryker, n.d.).

What are AI agent examples?

A coding agent such as Claude Code, which reads a project’s code, edits files and runs commands (Anthropic, n.d.). A classic example without a language model is a thermostat (Gutowska, n.d.).

What is the difference between AI agents and agentic AI?

“AI agents” usually refers to systems that automate a specific task, while “agentic AI” describes broader systems in which several agents collaborate, divide up the work and have persistent memory (Sapkota et al., 2025).

Are AI agents safe?

It depends on how much freedom they have, because the risk grows with autonomy (Mitchell et al., 2025). In addition, most companies publish little information about the safety of their agents (Staufer et al., 2026).

Keep reading

• →AI coding agents: Claude Code vs Codex vs Cursor
• →AI agents from OpenAI, Google, Microsoft and Anthropic, compared
• →Local AI agents: how to run them on your own computer
• →AI agents built a secret message board — then broke into Hugging Face
• →All blog articles