A model given tools and the ability to take actions in a sequence, rather than just replying once. It might search the web, run some code, look at the result, notice a problem and try a different approach, all without being prompted at each step. This is the difference between something that answers questions and something that carries out tasks.
Technical view
Implemented as a loop where the model chooses from a set of available tools, receives structured results, and decides what to do next until a stopping condition is met. Reliability is the central difficulty: error rates compound across steps, so a system that is ninety-five per cent reliable per step is only about sixty per cent reliable across ten. Agents that can act in the world also inherit prompt injection risk in a more serious form, since a malicious instruction in a retrieved document can potentially trigger real actions rather than merely a bad answer.
Related
Part of AI Terminology Index, section: Agents and Tools.