What is an AI agent?

What this video shows

Six words, typed twice: “A customer paid twice. Fix it.”

The first time they go to a chat. The chat answers at once with a short list: check the payment, refund the extra, email the customer. Good advice. Then the person opens the payment system, clicks, opens the mail, clicks again. Every step is theirs. The chat window sits dark the whole time.

The second time the same six words go to an agent. From here the person types nothing. The agent opens the payment system itself and finds two payments, the same size, side by side. It stops once, with one small card: “Refund the extra?” and an OK button. The person clicks OK. The agent sends the refund, reads the account again to be sure one payment is left, sends the email, and comes back with one word: Done.

I kept the picture the same in both halves on purpose. Same person, same window, same two systems. Only one thing changes: who moves.

The one stop is on purpose too. I do not want an agent to move money by itself, and I do not want to approve every click either. So in the small AI apps I build for daily work, the agent runs the routine part and stops at the steps you choose. Money is always one of them.

What you will learn

• The difference in one line: a chat talks, an agent does the work.
• The moves of an agent in plain words: it looks, it acts, it checks, it finishes, and it comes back to you. Anthropic’s guide for builders calls the same loop gather context, take action, verify work, repeat.
• Where you belong: at the money step, not at every step. Anthropic’s study of millions of real agent sessions puts it this way: oversight means being in a position to step in when it matters.
• How to tell a renamed chatbot from an agent: ask whether it can open your systems and act in them. If it can only answer, it is a chat. OpenAI’s guide for builders says it plainly: a simple chatbot is not an agent.

Where agents are, in numbers

Nothing with a digit is on screen in the video. The numbers belong here.

• 2025: fewer than 5% of enterprise applications had a task-specific agent (Gartner, August 2025). 23% of companies had scaled an agent anywhere in the business; 39% were still experimenting (McKinsey, November 2025).
• 2026: Gartner expects 40% of enterprise applications to include task-specific agents by the end of the year, and agentic AI spending to reach about USD 202 billion. Among companies with revenue over a billion dollars, 40% are scaling agents; smaller companies sit at 22% (McKinsey, August 2026).
• 2027: Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mostly for unclear value, cost, or weak controls. My guess for 2027: buyers learn to tell a renamed chatbot from a real agent. The ones that survive are small, bounded agents with a person at the money step. The agent in this video is that kind.

Questions people ask

Is a chatbot an AI agent?

No. A chatbot answers in text and stops. An agent can use your systems and keeps working until the job is done, checking its own work on the way. OpenAI’s guide for builders says it plainly: a simple chatbot is not an agent.

Why does the agent stop and ask?

Because moving money is a step a person should see. In the video the agent stops once, at the refund, and nowhere else. You decide which steps get a stop. Reading an account or writing an email does not need one.

What does “it checks” mean?

After the refund, the agent reads the account again and sees one payment left. It does not trust its own last step; it looks at something real. Anthropic’s builders’ guide calls this verifying the work, and it is the part that makes an agent safe to leave alone.

When is an agent the wrong tool?

When the job is the same steps every time, like sending the same invoice on the first of the month. Plain automation is cheaper and never surprises you. Agents are for jobs that need a look, a judgment, or an input that is never the same shape twice.

Further reading

• Building effective agents, Anthropic, December 2024. Workflows for fixed steps, agents when the path cannot be set in advance, and “the simplest solution possible”. The customer-support example with a refund is theirs.
• Building agents with the Claude Agent SDK, Anthropic, September 2025. The loop the video shows: gather context, take action, verify work, repeat.
• Measuring AI agent autonomy in practice, Anthropic, February 2026. Millions of real sessions; oversight as being in a position to step in when it matters.
• A practical guide to building agents, OpenAI, April 2025. What an agent is and is not, and human checks for high-risk actions such as refunds and payments.
• Introduction to Agents, Google, November 2025. The model as the brain, tools as the hands, and the levels from a plain model to a team of agents.
• ReAct: Synergizing Reasoning and Acting in Language Models, Yao and others, 2022. The research root of the look, act, check loop.
• Gartner press releases, June 2025 and August 2025, and McKinsey, “The state of AI”, November 2025 and August 2026. The numbers in the section above.