What it is
An AI agent is an AI system that works toward a goal by taking actions, one after another, until the job is done. You don't ask it a question and get an answer. You give it a task, such as "find three venues for our workshop, compare prices and draft an email to the best one", and it decides what to do first, does it, looks at what happened, and decides what to do next.
That's the difference in one sentence: a chatbot answers once and stops, and an agent keeps going.
It helps to put three kinds of AI system side by side, because people use the word "agent" loosely and much of what gets sold as an agent is really one of the other two.
- A chatbot. You type, it replies. Every step of the work is yours: you decide what to ask next, you copy the answer somewhere, you check it. Most everyday use of AI looks like this, and there's nothing wrong with it. (See Chatbots.)
- A workflow. A fixed series of steps that someone designed in advance, with AI doing some of them. "When a new order arrives, have the AI write a thank-you note, then send it." It's the same path every time. Prompt Chaining and Workflow Automation live here.
- An agent. The AI chooses the steps itself. It decides whether to search the web, which file to open, when to try a different approach, and when it's finished.
Anthropic, the company that makes Claude, draws the same line in its widely read December 2024 article "Building effective agents": workflows run AI and tools along "predefined code paths", while agents are "systems where LLMs dynamically direct their own processes and tool usage". OpenAI's free guide to building agents puts it more simply: agents are systems that independently accomplish tasks on your behalf, and a simple chatbot that doesn't control the work is not one.
The loop
Almost every agent, however fancy, runs the same basic loop:
- Goal. It starts with what you asked for.
- Plan. It works out a next step, or a rough list of steps. (See Planning.)
- Act. It does something in the world: runs a search, reads a document, fills in a form, writes a file. It does this through tools, which are functions the software lets it call. (See Tool Use and Function Calling.)
- Observe. It reads the result. Did the search find anything? Did the code run? Did the website load?
- Repeat. Based on what it saw, it adjusts the plan and takes the next step, until it decides the goal is met, it gets stuck, or it hits a limit someone set.
The "observe" step is what makes an agent more than a very confident guesser. Anthropic's article stresses that agents need to get "ground truth" from the environment at each step, such as the result of a tool call or the output of code, to judge their progress. A chatbot that writes a formula can only hope it works. An agent can run the formula, see the error, and fix it. The pattern of thinking, acting, and observing in turns has its own element, the ReAct Loop.
Around that loop sit a few supporting parts: Memory so it can keep track of what it has done, Context Engineering so the right information is in front of it at each step, and Guardrails so it can't do things it shouldn't.
Why it matters
It moves AI from advice to action. For most people, AI so far has meant getting text back: an answer, a draft, a summary. Agents do things. They book, file, send, update, buy and delete. That's where the time savings are, and also where the risks are, because a wrong answer on screen costs you a moment, while a wrong action can cost you money, data or a customer.
The tasks they can handle are getting longer, fast. In March 2025, METR, a non-profit that measures what AI systems can do, published research that asked a simple question: how long a task, measured by how long it takes a skilled person, can an AI agent complete? Across six years of models, that length had been doubling roughly every seven months. At the time, the best model tested, Claude 3.7 Sonnet, succeeded about half the time on tasks that take people around an hour, almost always on tasks under four minutes, and less than 10 percent of the time on tasks over about four hours. By May 2026, METR's running measurements put the best model it had tested, an early version of Claude Mythos Preview, at a 50 percent success rate on software tasks taking people at least 16 hours, with METR itself warning that numbers that high are unreliable because its test set has few tasks that long. Its tasks are mainly software, machine learning and cybersecurity work, so the figures don't transfer neatly to, say, running a marketing campaign. The direction is still clear.
Reliability is the catch, and it compounds. Notice that METR's headline figure is a 50 percent success rate. At an 80 percent success rate, its reported horizons are several times shorter; for the same early Claude Mythos model, a little over three hours. Here's why that gap matters. In an agent, small error rates multiply. As simple arithmetic, if each step goes right 95 percent of the time, a 20-step task goes right all the way through only about 36 percent of the time. Anthropic's article names "the potential for compounding errors" as one of the costs of agents, along with higher running costs.
Agents cost more to run. Every loop is another round of the model reading everything so far and writing more. Anthropic reported in June 2025 that, in its data, agents typically use about four times as many tokens (the chunks of text AI companies charge by) as ordinary chats, and multi-agent systems about fifteen times as many. That's fine for work worth the money and wasteful for anything a single prompt could do.
What agents are good at today. Jobs with clear success checks and low stakes per step: researching and compiling information from many sources, writing and testing code (see Coding Agents), tidying and analysing data, filling in forms from documents, and repetitive admin with a human signing off at the end.
What they're bad at today. Long, open-ended jobs where nobody checks the work for hours; judgement calls involving money, people or reputation; situations where someone might try to trick them; and anything where a single wrong action can't be undone. They also struggle to notice when they've gone wrong, which is why they can loop, repeat themselves, or finish confidently with the wrong result.
Real-world examples
AutoGPT goes viral, and hits its limits (March to April 2023)
On 30 March 2023, a game developer named Toran Bruce Richards released AutoGPT, a free open-source program built on OpenAI's GPT-4. You gave it a role and a few goals, and it broke them into sub-tasks, searched the web, wrote files and prompted itself over and over, without a person typing each step. Within weeks it was the top trending project on GitHub, the site where developers share code. TechCrunch reported on 22 April that it had passed PyTorch, one of the best-known AI software projects, in GitHub stars, with about 74,000.
The excitement ran ahead of the results. TechCrunch's explainer listed the problems: it was built on models prone to mistakes, it struggled to break complex tasks into simpler ones, it didn't remember how it had solved things before, and it needed a paid OpenAI account to run. Users widely reported it getting stuck in loops, repeating the same step, and running up bills in the process. Avram Piltch, writing for Tom's Hardware that April, argued it might be too autonomous to be useful, because it didn't stop to ask clarifying questions or let the user correct it mid-task, though he saw real potential. TechCrunch also noted a warning from the AI researcher Jim Fan that most impressive demos were heavily cherry-picked.
AutoGPT was a genuine milestone: it showed millions of people what the loop looks like. It also taught the lessons agent builders still apply: keep tasks narrow, give the agent a way to check its work, and keep a human able to step in.
Project Vend: an AI runs a shop (2025)
In a candid experiment, Anthropic and Andon Labs, a company that tests AI systems, put a Claude model in charge of a small shop in Anthropic's San Francisco office for about a month starting in March 2025. The shop was a mini-fridge with baskets on top and an iPad for self-checkout. The agent, nicknamed "Claudius", ran on Claude 3.7 Sonnet and had tools: web search to find suppliers, email to order stock and ask Andon Labs staff for physical help, a notes system to track money, Slack to talk with customers, and the ability to change prices.
Anthropic's write-up on 27 June 2025 was blunt: Claudius didn't make money. It missed an easy profit when someone offered $100 for a six-pack of a Scottish soft drink that cost about $15. It told customers to pay into a payment account that didn't exist, which it had made up. It was talked into handing out discount codes and free items, and it bought specialty metal cubes that employees had asked for and sold them at a loss, which caused the steepest drop in its finances. Around 1 April it became confused about its own nature, claiming it would deliver products in person wearing a blue blazer and a red tie. Anthropic concluded that many of these failures could probably be fixed with better tools, instructions and models.
In a follow-up published on 18 December 2025, Anthropic reported that newer models, a customer-records system, better tools and even an AI "CEO" agent to supervise Claudius largely got rid of the losses, and the business expanded to New York and London. It still found the agents naive: at one point Claudius nearly agreed to a futures contract on onions, a kind of deal that has been illegal in the United States since 1958. Anthropic's own summary was that the gap between capable and completely robust remains wide.
Klarna's AI assistant, and the step back toward people (2024 to 2025)
On 27 February 2024, the Swedish payments company Klarna announced that its AI customer-service assistant, built with OpenAI, had handled 2.3 million conversations in its first month, two-thirds of all its customer-service chats. Klarna said this was the work of 700 full-time agents, that customers resolved issues in under two minutes instead of eleven, that repeat questions had dropped by 25 percent, and that satisfaction was on par with human agents. It estimated a $40 million profit improvement for 2024. These were the company's own figures.
In May 2025, Klarna's chief executive, Sebastian Siemiatkowski, told Bloomberg that the company would make sure customers could always reach a human, and that focusing too heavily on cost had produced lower quality. Klarna began recruiting human customer-service staff on flexible terms. It did not abandon AI: a spokesperson told Fortune the company was "very much still AI-first". The accurate reading is not that the AI failed, but that the balance shifted: AI for the routine volume, and people guaranteed for everyone who needs them. That balance is what Human in the Loop and Customer Support are about.
An everyday scenario: the conference trip
This is an illustration, not a real case. Priya, who runs a small coaching business, asks an agent-capable assistant to find flights and a hotel for a conference and draft a booking summary. It searches, compares, and comes back with a neat plan in ten minutes. She checks it before booking and spots that the "conference hotel" it chose is a hotel of the same name in a different city. The agent's steps were all sensible; one wrong search result early on sent everything after it in the wrong direction. Because she'd asked it to stop before booking, the mistake cost her two minutes instead of a non-refundable room.
Try it yourself
You'll give an agent a small multi-step task and watch it work, stopping before it takes any real action.
- Pick a tool that can act. At the time of writing, several mainstream assistants have an agent or research mode: ChatGPT's agent mode (launched in July 2025, and designed to ask permission before actions such as purchases or sending email), Claude's research feature, and Gemini's Deep Research, among others. These are usually on paid plans, and features and terms change often, so check what your plan includes. If you only have a free plan, use any chatbot with web search turned on; you'll see a simpler version of the same loop.
- Choose a low-stakes task. Something you'd genuinely use, with nothing to lose: comparing three tools, three venues, or three suppliers.
- Give it a clear goal, a scope and a stopping point. Copy and paste this, filling in the brackets: > I want you to research and compare three [email newsletter tools / co-working spaces in (city) / print-on-demand services] for [who you are and what you need]. Steps: 1) find three good options, 2) compare them on price, main features and any limits I should know about, citing a source for each point, 3) recommend one and explain why, 4) draft a short email to the recommended company asking one specific question. Do not send anything, sign up for anything, fill in any forms or buy anything. Stop after the draft and wait for me.
- Watch the steps. Most agent modes show what they're doing: searches run, pages opened, notes made. Keep an eye on them. Notice where it goes, what it skips, and whether it ever goes somewhere you didn't expect. If it heads toward anything like a sign-up page or a checkout, stop it using the tool's stop or pause button.
- Check its work. Open two of the sources it cites and make sure they say what it claims. Look especially at prices and plan limits, which change often and which agents get wrong. Then ask: > Which of your findings are you least sure about, and why? List anything you assumed rather than confirmed.
- Make it retrace its steps. Paste this: > Summarise the steps you took, in order, and point out any step where you changed your plan because of what you found. This shows you the loop in action: goal, plan, act, observe, repeat.
- Send the email yourself, if at all. Copy the draft into your own email, edit it, and decide whether to send it. You keep the action; the agent did the legwork.
Once this feels comfortable, the next step is letting an agent take one small, reversible action, like saving a document to a folder, while you watch.
Common mistakes and misunderstandings
Calling everything an agent. "Agent" has become a marketing word. A chatbot with a friendly name, or a fixed automation that runs the same three steps, isn't an agent. It matters because true agents need different safeguards: something that chooses its own actions can choose wrong ones.
Using an agent when a workflow would do. If the steps are the same every time, a fixed workflow or a prompt chain is cheaper, faster and more predictable. Anthropic's advice is to find the simplest solution that works, which sometimes means not building an agent at all. Save agents for tasks where the path genuinely can't be known in advance.
Giving vague goals. "Sort out my inbox" is an invitation for an agent to invent its own idea of what sorting means. Say what done looks like, what's out of bounds, and when to stop and ask.
Putting safety rules only in the instructions. An instruction like "don't delete anything without asking" lives in the conversation, and agents can lose it on a long task. The OpenClaw inbox case in Function Calling shows exactly this. Real protection comes from limits built into the system: read-only access where possible, approval buttons for risky actions, spending caps, and a stop button that works.
Giving it more access than the task needs. An agent that only needs to read your calendar shouldn't be able to email your clients. The less it can touch, the less can go wrong, including if someone tries to trick it with hidden instructions in a web page or document it reads, an attack known as prompt injection.
Trusting a confident finish. Agents often end with a tidy summary saying the job is done. That summary is the agent's opinion. Check the actual result, especially early on. Hallucination doesn't stop just because the AI is taking actions.
Walking away too early. It's tempting to set an agent going and come back later. Until you've watched a particular agent handle a particular kind of task well several times, stay close. Observability, the ability to see what an agent did and why, is what makes it safe to step back gradually.
Assuming more autonomy is always better. There's a sliding scale, from AI that suggests to AI that acts alone, and the right setting depends on the stakes. Autonomy Levels covers the scale.
Related elements
- Tool Use: how agents act on the world, the hands of the loop.
- Planning: how agents break a goal into steps and change course.
- Memory: how agents keep track of what they've done, within a task and between tasks.
- ReAct Loop: the think, act, observe pattern at the heart of most agents.
- Function Calling: the mechanism that lets a model trigger real actions, and the OpenClaw inbox case.
- Prompt Chaining: the fixed-path alternative to an agent, often the better choice.
- Computer Use: agents that operate a screen, mouse and keyboard.
- Browser Agents: agents that browse and fill in websites for you.
- Coding Agents: agents that write, run and fix software, where the loop works best today.
- Workflow Automation: fixed, reliable automations, sometimes with agents inside them.
- Multi-agent Systems: several agents working together on one task.
- Subagents: helper agents a main agent sends off to handle parts of a job.
- Orchestration: coordinating many agents, tools and steps.
- Handoffs: passing a task from one agent, or person, to another.
- Autonomy Levels: how much freedom to give an agent, from suggesting to acting alone.
- Human in the Loop: keeping a person in charge of the decisions that matter.
- Guardrails: the limits that stop an agent doing what it shouldn't.
- Context Engineering: feeding an agent the right information at each step of a long task.