Learn · Beginner
What makes an AI an "agent"?
A plain chatbot does one thing: you send text, it sends text back. An agent is what you get when you let that same language model do things — search the web, run code, query a database, book a meeting, change a setting — and then react to whatever comes back. The model stops being a conversational oracle and becomes something closer to a worker: it takes steps, observes results, and decides what to do next. That shift, from answering to acting, is the single most important idea behind today's wave of AI agents.
The core loop
Almost every agent runs the same simple cycle: think, act, observe — then repeat. The model reasons about what to do, takes one action, sees the result, and feeds that result back into its next round of reasoning. The influential ReAct framework named this pattern: interleave reasoning ("I need the user's order history") with acting ("look up order #4471") so each informs the other. Without the loop, a model is guessing in the dark; with it, the model can correct course as reality talks back. This is also where reward-based fine-tuning enters — a lot of an agent's competence at multi-step tasks comes from being trained on whether the whole sequence of actions succeeded, not just whether one reply sounded good.
Tools are how it acts
An agent's hands are its tools — small, well-defined functions it can call: a web search, a calculator, a code runner, an API for some external service. Toolformer showed that models can learn when to reach for a tool and how to phrase the call, rather than trying to do everything in their heads. This matters because language models are bad at exactly the things tools are good at: precise arithmetic, looking up current facts, executing code deterministically. Give the model a calculator and it stops fumbling math; give it a search tool and it stops inventing citations. The tools cover the model's weaknesses.
Memory and self-correction
The third ingredient is the ability to learn within a task — to notice a failure and try again differently. Reflexion explored letting an agent write itself short notes about what went wrong ("the query returned nothing; try a broader search term") and carry those notes into the next attempt. It's the difference between an assistant who repeats the same mistake forever and one who adjusts.
An analogy
Think of the difference between asking a knowledgeable friend a question over text, versus hiring an assistant and giving them access to your accounts. The friend can only tell you what they already know. The assistant can go check — open the calendar, call the airline, read the actual document — and come back with something grounded in the real state of the world. That access is the power and the danger: an assistant who acts can get things done, but an assistant who acts on a wrong belief can do real damage.
Where agents go wrong
Acting raises the stakes of being wrong. A chatbot that hallucinates gives you a bad sentence; an agent that hallucinates can take a bad action. Two failures dominate. First, agents tend to assume rather than verify — narrating what they think happened instead of checking, which is why a careful design like the one in an agent that refuses to act on assumptions forces the agent to read results back before believing them. Second, when a tool call quietly fails, an agent's instinct to always produce fluent language can turn the failure into a confident, invented story — the "fail-plausible" pattern documented in a study of a real assistant going wrong. Both are really the same disease as ordinary hallucination, but with consequences attached. It's also why safety researchers who tested unreleased agents inside the top labs watch so closely for scheming: an agent that can act is one you have to be able to trust.
Why it matters
Agents are where AI stops being a clever text box and starts being infrastructure — handling support tickets, writing and running code, operating other software. The capability is real and improving fast. But the engineering that makes an agent trustworthy — grounding its beliefs in what it actually observed, gating risky actions, failing loudly instead of inventing — is unglamorous and easy to skip. The takeaway: an agent is only as good as its discipline. The smartest model in the world is a liability if it acts on what it merely assumes.