AI 101 — What is an AI agent?

An AI agent is a system that uses a language model to do a job for you — not just answer a question, but pick the next step, use tools, and keep going until the work is finished.

A chatbot waits for you. An agent takes the goal and acts. That is the whole difference, and it is why the word is on every product page this year.

Why it matters right now

Software used to automate work by following a script a human wrote in advance. Agents flip that: the model writes the next step as it goes. OpenAI's working definition is blunt — agents are systems that independently accomplish tasks on your behalf. Anthropic draws a sharper line: a workflow is a model walked down a path you already coded; an agent is a model that chooses its own path and its own tools.

That distinction stopped being academic this week. Cursor launched Origin, a code host built for agents that do not wait in a human review queue. Researchers used an agent to port a 250,000-line Fortran weather model onto GPUs and ran a real typhoon simulation 5.1 times faster, after checking every kernel against the original. The products are shipping. The vocabulary has not caught up.

The mental model

An agent is not a new kind of brain. It is a familiar model, wrapped in three parts.

The model is the decision-maker. It reads your goal, looks at what just happened, and chooses the next move — often by thinking the problem through in chain of thought before it acts.

The tools are the hands. Search, a calendar, a code editor, a browser, a database. The model cannot do any of those things on its own; it asks a tool, gets a result, and folds that result back into its working memory. MCP is the socket those tools plug into.

The loop is what makes it an agent. Do something. Look at the result. Decide the next thing. Repeat until the job is done, a limit is hit, or the system hands control back to you. Anthropic's engineers have described the pattern in one line: models using tools in a loop.

That loop is also why agents cost more and fail louder than a single chat reply. Every extra step is another chance to drift. Good agents check the world after each action — a test that passed, a file that actually changed, a search result that exists — instead of trusting their own plan.

The everyday analogy

A chatbot is a well-read librarian. You ask a question; they point you to the right shelf, or write you a summary. You still have to walk over, pull the books, and do the work.

An agent is a research assistant you send into the stacks. You say, "find every mention of this company in last year's reports and draft a briefing." They walk the aisles, pull volumes, notice a gap, go back for another book, and come back with the document. Same building, same books. The difference is they left the desk.

You still decided the job. They decided the steps.

Common misconceptions

"An agent is a new kind of model." No. The model is usually the same one that already answers your questions. Agency is the system around it — tools, a loop, and a stop condition. Swap the model and you have a faster or slower assistant. Remove the loop and you have a chatbot again.

"Agents work without humans." They still need a goal, a set of allowed tools, and usually a reason to stop. Many production agents pause for approval before they send an email or merge a change. Anthropic's own advice is to start simple and only add autonomy when a fixed workflow is not enough, because agents trade cost and latency for flexibility, and errors compound.

"More agents are always better." A Google Research study tested 180 agent setups and found the opposite of a free lunch. Multi-agent teams helped a lot on work that splits cleanly into parallel pieces. On sequential work — where step two depends on step one — extra agents often made results worse. Their model picked the right architecture for 87 percent of unseen tasks. Headcount is not a strategy.

"If it can use tools, it is safe." Tools are hands. Anything the agent reads — a web page, a PDF, a ticket — can try to talk those hands into the wrong job. That is prompt injection, and it gets more dangerous the more the agent is allowed to do.

Where to learn more

Anthropic's "Building effective agents" essay is the clearest map of the pattern, including when not to build one. OpenAI's practical guide is the product-team companion: what an agent is, when the extra complexity is worth it, and how to put up guardrails. IBM's explainer is the gentlest walkthrough of the loop. And if you want the research that punctures the "just add more agents" slogan, start with Google's scaling paper.

Related reading: What is MCP? · What is prompt injection? · What is chain of thought?

Which job would you actually hand to an AI agent first — and which one would you never let it touch? Tell us in the comments.

Sources: OpenAI — A practical guide to building agents · Anthropic — Building effective agents · IBM — What are AI agents? · Google Research — Towards a science of scaling agent systems · Cursor — Origin code hosting · arXiv — agent ports a 250k-line weather model