What is an AI agent? Definition, levels of autonomy, examples, and limitations
Un AI agent An agent is a piece of software that receives a goal, decides on the steps to achieve it, uses tools to act (read files, search the web, send a message, complete an application), observes the result of each action, and repeats until the task is completed or requires human intervention. This is the key difference with a chatbot: a chatbot answers, an agent performs. By 2026, the word is everywhere, from Copilot to ChatGPT, including Claude, Gemini, and Le Chat, and it refers to very different things, from a simple configured assistant to a system that works independently for twenty minutes. This article provides a sound definition, explains how an agent functions, proposes a scale of autonomy to help understand them, and frankly states what an agent does poorly and under what conditions it can be entrusted with tasks. For the enterprise-level version, including development environments and business use cases, see our AI Agents Guide.
Article published in January 2025 and completely rewritten in September 2026. The first version included the general definition of autonomous agents; this one compares it to what agents have become in everyday tools, adds a scale of autonomy, corrects a misconception about learning and details the rules of supervision.
A definition that holds true, and what it excludes
The academic definition predates major language models: in Stuart Russell and Peter Norvig's reference manual, an agent is any system that perceives its environment through sensors and acts on that environment through effectors, pursuing a specific objective. Applied to the enterprise of 2026, this definition remains accurate, provided the terminology is translated. The environment includes your files, emails, applications, and the web. The sensors and effectors are... tools The model has access to these: search, document reading, browser, email, calendar, databases, and connectors to business software. The brain is a language model that reads the objective, reasons, chooses the tool, interprets the result, and decides what to do next.
Anthropic introduced a distinction at the end of 2024 that has become the benchmark for those who build these systems. On one side, the workflows, where the model and tools are linked according to a path pre-written by a human. On the other hand, the agentswhere the model itself directs its process and the use of its tools, step by step, based on what it observes. The difference lies not in the power of the model but in who decides on the stepsIn a workflow, it's the designer; in an agent, it's the model, at the time of execution.
This definition therefore excludes three things that are nevertheless often called "agents". One chatbot, even excellent, who responds to every message without doing anything else. A personal assistant (a GPT in ChatGPT, a declarative agent created with Copilot's Agent Builder, a Claude project), which is a chatbot equipped with instructions and documents, valuable but without an action loop. And a classic automation (Power Automate, Zapier, Make, a script), each step of which has been written in advance, even if one of the steps calls for a language model. These are not lower degrees of quality: they are different objects, often preferable, as we will see below.
How an agent works: the loop
Let's take a simple request: "Prepare tomorrow's meeting with the client Durand." A chatbot would respond with generic meeting preparation tips. A human agent, on the other hand, starts by checking the calendar to find the meeting and its participants. They then search for recent email exchanges with this client, open the minutes of the previous meeting in the shared folder, note the commitments made by both parties, and check the status of the quote in the CRM. At each step, they read what they've found and decide on the next action: if the minutes mention a dispute over an invoice, they'll retrieve the invoice. Finally, they draft a one-page memo with the proposed agenda, key points to consider, and questions to ask, then stop and submit it to you. If they had to send an email to the client, a well-trained agent would have asked for your approval first.
This loop, objective, plan, action using a tool, observation of the result, adjustmentIt runs until the agent deems the work complete, encounters an obstacle, or reaches a predetermined limit (number of steps, budget, prohibited action). Five components are necessary to run it. One language model capable of reasoning and reliably calling up tools, which is not the case for all models or sizes. tools and connectors which provide access to data and applications; the standard MCP In two years, it has become the universal interface for connecting an agent to software. working memoryThat is to say, the context accumulated throughout the mission, sometimes supplemented by persistent memory between missions. instructions and safeguards : the scope, the authorized actions, those requiring validation, the expected tone. And a monitoring interface where the human sees what the agent is doing, approves sensitive actions and regains control.
One point needs correcting, because the first version of this article stated it and many presentations reiterate it: an agent does not learn from its mistakes. The model that drives it does not change during use. What can be improved are its instructions, its tools, its memory, and the examples it is given, and it is a human who evolves these based on observation. An agent that makes a mistake on Monday will make the same mistake on Tuesday if no one has changed its configuration. This limitation is also a safeguard: its behavior is reproducible and auditable.
Five levels of autonomy to help you find your way around
The word "agent" in 2026 covers a whole spectrum, and most misunderstandings arise from the fact that different levels are discussed without explicitly stating this. Here is a simple gradation, based on a single question: who decides on the steps, and who verifies them?
| Level | Who decides on the steps, who checks | Examples in 2026 |
|---|---|---|
| 0.Chatbot | You respond to every message; no action is taken. | Copilot Chat, ChatGPT, Claude, Gemini, The Chat in Conversation |
| 1. Equipped Assistant | You request one action at a time, the model executes it. | "Search the web," "read this attachment," a GPT or a Copilot reporting agent with their documents |
| 2. Automated Workflow | A human wrote the steps; the model fills in each box | Power Automate or Make with a call to a model, a Copilot Studio agent with a fixed scenario |
| 3. Supervised Agent | The model chooses the steps; a human validates the sensitive actions and the result. | Claude CoworkCopilot CoworkChatGPT in agent mode, Deep Research, Claude Code, Copilot Studio agents with shares |
| 4. Autonomous Agent | The model decides and acts alone; humans perform checks afterward, on a sample. | Automatic ticket sorting and response, monitoring and alerting, agents triggered by an event |
Two points to note. First, a single product often covers several levels: "agent" in Microsoft 365 Copilot refers to both a level 1 declarative assistant and a level 3 or 4 Copilot Studio agent with actions and triggers; ChatGPT is a conversational chatbot and a supervised agent in agent mode or Deep Search. Second, the most useful level is not necessarily the highest. The vast majority of value created in teams today is at level 3, where humans retain control over the actions that matter. Level 4 is reserved for tasks where errors are inexpensive, reversible, and detectable. Our article on Four ways to delegate work to AI details what each level changes for the person delegating.
What does an agent look like in the tools of 2026?
Two years ago, building an agent required a developer, a framework, and weeks of work. Today, level 3 agents are delivered within the tools that teams already have. Microsoft 365 CopilotThe Researcher and Analyst agents conduct multi-step data searches or analyses on company documents; Copilot Studio allows the creation of agents that operate within business applications, and the mode Cowork assigns a complete task to the user's files. In ChatGPTThe agent mode navigates the web, fills out forms and produces files; Deep Research performs dozens of searches to deliver a sourced report. Claude supports Cowork, who works on your files and connected applications up to the deliverable, and Claude Code for the developers. Gemini et The Chat Mistral's agents offer their own search and execution agents. We compare these environments in our Claude comparison CoworkCopilot Cowork and ChatGPT Work.
The common thread among all these products, and the real innovation compared to 2024, is that the agent sees what you see: your desktop, your files, your email, your applications, through standardized connectors. An agent with access only to its general knowledge cannot work for you; an agent connected to your environment can, and that is precisely why it needs monitoring. The second innovation is the model selection : publishers now offer models of different sizes, and an agent who makes fifty calls in a row does not need the largest one for each of them.
Three concrete examples, from the simplest to the most ambitious
The personal agent. Every Friday, a sales manager asks her agent to prepare her weekly review: he rereads emails and meeting minutes from the past five days, extracts commitments made, checks the CRM for deals that haven't progressed, and writes a two-page memo with follow-up notes. Nothing is sent without her input. The difference? An hour of preparation reduced to ten minutes of review, and most importantly, follow-ups that are no longer missed. This is the ideal use case to start with, because mistakes are cost-free and the user is in the best position to judge the outcome.
The team agent. A generic inbox receives around a hundred requests per day, a mix of customer questions, complaints, internal requests, and spam. An agent reads each message, categorizes it, locates the relevant case, prepares a draft response from the knowledge base, and creates the ticket in the tracking tool. For the first few weeks, a human approves each response before it is sent; subsequently, only complaints and cases flagged as uncertain by the agent are reviewed. This gradual transition from level 3 to level 4, category by category, is the sensible way to automate the process.
The integrated business agent. A tendering department receives a 300-page tender document. An agent extracts the award criteria, required documents, timeline, and potential obstacles, compares them to the company's references and past responses, and produces a reasoned go/no-go grid for the committee. Here, the agent makes no decision; they transform two days of reading into half a day of discussion. Examples of building agents for business workflows and the case of the HR function orchestrated by agents extend this logic.
What an agent does wrong, and why it matters
The mistakes are piling up. A model that is 95% reliable in one step becomes only 60% reliable in a ten-step process, and 36% reliable in a twenty-step process, because each step relies on the previous one. This is why agents excel in short, well-defined tasks and underperform in long processes where no one checks the intermediate results. Breaking down the task, imposing checkpoints, and giving the agent the means to verify their own work (reviewing, testing, comparing to a source) makes all the difference.
He obeys what he reads. An agent that reads emails, web pages, or documents can find hidden instructions written specifically for it and follow them as if they came from you: this is prompt injection, the primary vulnerability of language model-based systems. An agent that has the right to send emails and reads its inbox is, by definition, a target. Hence, two rules: limit irreversible actions and treat all content read by the agent as data, never as an instruction.
It costs more than a conversation. An agent's task consumes ten to one hundred times more computing power than a question asked to a chatbot, because it involves dozens of calls and rereading its entire context each time. On a team scale, this is reflected in the bill. And the most underestimated cost isn't there: it's supervision and verification time, which never completely disappears and must be taken into account in the profitability calculation.
The market is exaggerating. In June 2025, Gartner estimated that over 40% of agentic AI projects would be abandoned by the end of 2027 for three reasons: spiraling costs, unclear business value, and insufficient risk controls. The same firm denounced "agent washing"—the rebranding of chatbots and traditional automation as agents—and counted only about 130 vendors offering genuine agents among the thousands claiming to do so. The tiered framework above is a good countermeasure: ask any vendor who decides on the steps and who verifies them.
When to use an agent, and when to do without one
The most useful rule comes from the teams building them: start with the simplest solution that works, and only add autonomy when it delivers something measurable. If the task is repeatable and stableWith consistently identical steps, a classic workflow (level 2) will be cheaper, faster, and more predictable than an agent; the agent would add randomness without providing any benefit. If the task requires judgment at each stageGiven that data is scattered across multiple tools and the path varies depending on what is found, a supervised agent (level 3) is the right tool: this is the case for meeting preparation, file analysis, and in-depth research. If the error is costly or irreversible (a payment, a contractual commitment, an external communication, a decision that affects a person), it is a human who acts, with an assistant who prepares; the agent can instruct, not decide.
Finally, there is one case where several agents are better than one: when the mission involves a mix of distinct skills (research, writing, verification, calculation), the roles are distributed among specialized agents, coordinated by an orchestrator. This architecture reduces errors and costs, provided it is not deployed before it is needed; we explain this in our article on multi-agent systems.
The supervision rules that make an agent usable
An agent is not an autonomous collaborator who can be trusted based on their good looks; it is a system that is managed. Organizations that benefit from it apply, in one form or another, the same rules.
- A written perimeter : what data the agent can read, what tools it can use, with the lowest possible rights, never those of an administrator.
- Actions are reversible by default : the agent prepares, proposes, drafts; sending, paying, deleting and publishing require human validation.
- A journal of actions : every tool called, every document read, every decision made is tracked, in order to understand an error and respond to an audit.
- Real-world case tests before commissioningwith difficult examples and some pitfalls, then indicators tracked over time: rate of human intervention, errors detected, time actually saved.
- An official appointed by agent, as with any tool that acts on behalf of the company, and a usage charter known to the teams.
These rules align with what the European AI regulation is gradually requiring of companies, particularly transparency and human oversight of systems that affect individuals. Our article on mastering agent-based AI without hindering innovation and our AI systems governance kit provide the documents to implement them.
Where to start
The best first agent is the one who, under supervision, performs a task you already do weekly and whose results you can assess: an activity review, a case summary, monitoring, or preparing for a meeting. You give them access only to the necessary files, you write down the objective as you would for a new colleague (the expected result, the sources, what they shouldn't do), you proofread everything the first time, then less and less frequently. After a month, you know their worth, their cost, and whether the task is worth assigning to a team member. This path, from individual to collective to professional, is the one followed by organizations where agents deliver on their promises; the opposite, starting with a grand project for autonomous agents, contributes to high dropout rates.
F.A.Q
Is an AI agent the same thing as a chatbot? No. A chatbot responds to your messages; an agent receives an objective, chooses the steps themselves, uses tools to act, and verifies the result. The same product can be either one depending on the method used.
What is the difference between an AI agent and an automation (RPA, Power Automate, Zapier)? In automation, a human has written each step in advance; in an agent, the model decides the steps at runtime. Automation is more predictable and less expensive for stable tasks; the agent handles variable cases that require judgment.
Do you need to know how to code to create an AI agent? No for levels 1 to 3: Copilot Studio, GPTs (which OpenAI announced will be withdrawn by December 2026), projects and Cowork Claude's or ChatGPT's agent mode can be configured using natural language. Code is still required to integrate an agent into a business system or for autonomous agents triggered by events.
Does an AI agent learn from its mistakes? Not by itself: the model does not change during use. It is its instructions, its tools, its memory, and its examples that a human evolves. This is a limitation, but also a guarantee of reproducible behavior.
Can an AI agent work without supervision? Only for tasks where errors are inexpensive, reversible, and easy to detect, after a period of supervised operation and sample control. For anything that involves the company or affects an individual, a qualified human being is required.
How much does an AI agent cost? Three factors are involved: the tool license, computing power (an agent's mission costs significantly more than a conversation), and the human time required for supervision and verification. It is this last factor that determines profitability.
Sources. Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, for the definition of the agent; Anthropic, Building effective agentsDecember 2024, for the distinction between workflows and agents and the "start simple" rule; Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027June 25, 2025, for dropout figures and "agent washing".
To learn more: our article When AI becomes a colleague, the article Why 2026 is the year of AI agents, our analysis of the rise of AI agents and governance challenges and the explanation of Deep Research.