What Is an AI Agent? (2026)

You are reading Chapter 25 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

An AI agent is a system you give a goal rather than a question: it picks its own steps, uses tools to take them, checks what came back and keeps going until it decides it is finished. The difference from the assistant you have used so far is not intelligence. It is that the output stops being a draft and becomes something that has already happened.

This chapter opens AI 303, the third of the course's three tracks: AI 101 gets you running, AI 202 teaches you to prompt properly, and AI 303 builds work that runs while you are not watching. It gives you a plain definition, the arithmetic that decides whether a multi-step job finishes clean, the measured evidence on what agents complete, and a four-question test for which of your own work can run unwatched. It assumes you have the prompt playbook from Chapter 24, so if you skipped it, start there.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

For 24 chapters, everything the assistant produced was text on a screen. You read it, you judged it, and only then did anything real happen. An agent takes the step itself.

An assistant replies. One turn: your question in, text out. The output is a draft every time, and a mistake costs you the minute it took to read it. Every step of the actual work is still yours, and nothing happens while you are asleep.

A system acts. You give it a goal, and it takes many turns, most of them without you watching. The output is something that has already happened, and a mistake costs whatever the action cost. The work continues while your attention is elsewhere, and you are accountable for actions you never read.

Anthropic's engineering team draws the useful line. A workflow runs through code paths somebody wrote in advance. An agent directs its own process and its own use of tools. One follows a route you fixed, the other picks the route.

That decides how each one fails. A workflow can only fail in ways you could list beforehand. An agent can fail in ways nobody listed, including by doing something sensible that you would never have chosen. Many products sold as agents are workflows, and that is often the better buy: cheaper, faster, and far easier to diagnose when something goes wrong. If you want to see what building either one involves, our guide to building AI agents covers the tools.

Many tools use the word agent loosely. In June 2025 Gartner noted a lot of relabelling of assistants, automation tools and chatbots as agents. So judge each one by two questions instead. What is it allowed to touch, and what happens if it is wrong?

Those two answers matter because of what changes when it acts. A draft sits on your screen until you approve it. An action is an email sent, a record changed or a file saved, so it pays to know which kind of work you are handing over.

Before you read on, write down one job you would hand over if you trusted it completely. The rest of this chapter is about whether you should.

This is the most useful idea in the chapter, and it is arithmetic rather than opinion. A system that is 95% reliable at each step is not a 95% reliable system. Chain the steps and the chance that every one goes right is 0.95 multiplied by itself once per step:

Put it against a real job. Say your Monday report takes 8 steps: pull this week's numbers, pull last week's file, match the line items, find the 3 biggest movers, look up why the largest one moved, write the commentary, format it, send it to 6 people. At 95% a step, you get a clean report on 66 Mondays out of 100. Cut it to 4 steps and you get a clean report on 81 Mondays out of 100, which is why shorter jobs work best.

People expect errors to average out. They feed forward instead. A figure fetched wrongly at step 2 is not questioned at step 5, it is written up in confident prose at step 6 and emailed at step 8. The system never flags that the figure looked odd, because it has no way to notice. This is the same untroubled confidence Chapter 8 taught you to catch, with more steps stacked on top and none of them read by you.

A demo is 3 steps long. Your real job is 15. The demo was honest and still told you almost nothing about how your job would run.

You can see the effect in a measured setting. A May 2025 study from Salesforce AI Research tested agents on realistic CRM work. The leading agents got around 58% of tasks right when the work came as a single request, and approximately 35% when it played out over several turns. The kind of work stayed the same. The exchange got longer.

Two rules follow, and both shape the rest of this track. Make the chain shorter, because every step you remove is a multiplication you no longer pay for. And make each step fail loudly, so a run that has gone wrong stops instead of finishing as if nothing happened. Chapter 16 taught you to break a big job into steps, and Chapters 28 and 34 build the checks that stop a bad run.

What predicts whether an automated job will work is not how clever it is. It is how many steps it takes and how visible each one is when it breaks.

You have probably been pitched agents several times, so here is the version with numbers attached. The short answer: they are already strong on short, well-defined jobs, and the length of job they can handle keeps growing.

The most useful finding in the field is about how long a task is. In March 2025 METR measured tasks by how long they take a skilled person, then checked how often models finished them. The models of that time succeeded almost 100% of the time on tasks taking a person under 4 minutes, and under 10% of the time on tasks taking more than around 4 hours. The length of task they could handle had been doubling roughly every 7 months.

By June 2026, METR's estimate of the task length one frontier model, GPT-5.6 Sol, could finish half the time had reached around 11.3 hours. In the same report METR said it did not consider any of its numbers a robust measurement, because the result depended heavily on how it treated the model's attempts to cheat the tests. Read the trend with that caveat attached.

Whole projects handed over end to end are a different matter. An October 2025 study from the Center for AI Safety and Scale AI, the Remote Labor Index, gave agents 240 real freelance projects across 23 kinds of digital work, together representing over 6,000 hours of human work. The best agent fully automated 2.5% of them. So hand an agent the smaller pieces of your week, not a whole job.

Company-wide programmes have a harder time, and the reason is worth knowing. A July 2025 report from MIT's NANDA project found the gap was not model quality. It was tools that do not learn, integrate poorly or do not match the way people actually work.

The Salesforce study adds one detail worth carrying into Chapter 27. The agents showed little built-in sense of what to keep confidential. Targeted instructions helped, sometimes at a cost to task performance, which is why Chapter 27 covers the settings that handle it.

This is good news for you, and it is why this track is written for one person rather than a procurement team. Almost everything that makes a company programme fail, the integration, the change management, the committee, is absent when one person automates one job they already do. What works today is bounded, repeatable, checkable and cheap to run again: gathering things, sorting them, pulling material together, drafting from what was just retrieved, routing what arrives. That covers most of the recurring work in your week.

Two questions decide this, and neither is about how capable the tool is. Can you undo it, and would you notice if it went wrong? Run each step of a job through these four checks, in order:

Notice what the test never asks: how hard the task is. Reversibility and visibility decide where the line sits.

A draft sitting in your outbox, a summary written into a document, a file placed in a folder, a list of what arrived overnight with the important items marked. If the worst case is that you open it and find it useless, let it run. You lose a few minutes on a bad morning and get back hours across a month, and Chapter 28 has you build exactly one of these.

Sending, publishing, paying, deleting, changing a record somebody else relies on, anything a customer or a colleague sees before you do. Not because it will go wrong on any given run. It very likely will not. But when it does, the cost bears no relation to the work you saved, and that is a bad trade at any completion rate in this chapter.

A job that breaks loudly is safe. A job that breaks quietly is the dangerous kind, because you keep trusting it. The weekly digest that silently stopped reading one of its 5 sources in March is worse than the one that crashed in March, because you find out in July. When you design one, design how it reports failure at the same time as the work it does.

When it acts on your behalf, your responsibility does not move. Your name is on the message it sent, and saying you had not read it has never worked with a manager or a customer. What you are buying is a shift in where your attention goes: less time doing the steps, more time specifying the job and checking what came out. That is a gain when the ratio is right, and a loss when you automate something you never needed done.

Before this chapter, every output was a draft and every mistake cost you reading time. After it, you can define an agent without the word autonomous, work out why a 20-step job at 95% a step goes wrong roughly 2 times in 3, and say which of your own steps are safe to run unwatched.

You carry two things into Chapter 26: your weekly job written out as numbered steps, and the reversible or not mark against each one. Chapter 26 starts connecting the assistant to the drive, mail, chat and notes where your work already sits, and it needs to know which of your tools those steps touch. For a look at one agent product before then, our breakdown of ChatGPT Work walks through OpenAI's version. More guides like this one are in the AI Central Library.

An AI agent is a system you give a goal rather than a question. It chooses its own steps, uses tools to take them, checks the results and decides for itself when it is finished. An assistant hands you a draft, while an agent hands you something that has already happened.

A workflow follows a path somebody fixed in advance, so it can only fail in ways you could list beforehand. An agent chooses its own path, so it can fail in ways nobody listed. Many products sold as agents are workflows, which are often cheaper and easier to fix.

Per-step reliability multiplies. At 95% per step, a 5-step job finishes clean 77 times in 100, a 10-step job 60 times and a 20-step job 36 times. Errors also feed forward: a wrong figure at step 2 is treated as fact by every later step.

They are strongest on short, well-defined jobs. In a May 2025 Salesforce study, leading agents managed around 58% of CRM tasks asked as a single request and approximately 35% across several turns. In an October 2025 study of 240 real freelance projects, the best agent fully automated 2.5% of them. Keep the jobs you hand over short and easy to check.

Work that only reads, gathers, sorts or drafts, and that you could undo in under a minute. Anything that sends, publishes, pays or deletes, or that a customer or colleague would see first, should wait for your approval.