What is an AI agent?
The step past chat: a model that plans, uses tools, and keeps going until the job is done.
An AI agent is a model arranged to act rather than just answer. Instead of returning one block of text, it works in a loop: it plans, calls a tool such as a search or a file reader, reads the result, and decides what to do next. The intelligence comes from the language model. The agency comes from the harness that lets its words trigger actions and feeds the results back in.
A loop, not a single answer.
Ordinary chat is one pass of inference: question in, answer out, done. An agent inserts a middle. After answering, the model's output can name an action, a piece of software notices, runs it, and hands the result back as more text for the model to read. The model then decides whether the job is finished or another step is needed. Each turn around the loop is still ordinary inference; what makes it agency is that the model's own words choose the next move, instead of a human typing it.
What the model needs to act.
Three things turn a model into an agent. Tools it may call, described precisely enough that the model knows when each one fits. A set of standing instructions that say what the job is and what is out of bounds, which is the work of a system prompt. And a stopping rule, because a loop that never ends is worse than no loop at all. None of these change what the model knows; they change what its instructions can cause to happen, which is where the engineering effort actually goes.
Where agents meet your documents.
The most useful research agents are grounded ones: at each step, the model works from retrieved passages instead of what it happens to remember, and it cites what it used. That constraint matters more as autonomy grows, because a free-running loop has more chances to invent confident details. Running such an agent on your own machine adds a quiet benefit: the actions, the files it reads and the reasoning all stay local, so the more capable the loop becomes, the less you have to trade privacy for it.