AI Agent

The defining feature of an agent is not intelligence but control flow. In a plain completion, the application decides what happens next; in an agent, the model does. Each turn of the loop appends the previous action and its observed result to the context, so the model is reasoning over a growing transcript of what it has already tried. Tools are exposed as named functions with argument schemas, and the loop terminates when the model emits a final answer, hits a step limit, or trips a guard the application defines.

That inversion of control is what makes agents useful and what makes them hard. Useful, because the number of steps needed does not have to be known in advance — the agent can look something up, discover it was the wrong thing, and try a different query. Hard, because every additional step multiplies the ways a run can drift: a misread tool result at step three quietly poisons the reasoning at step nine, and the failure surfaces only at the end.

The most common misconception is that an agent is a model with a better prompt. It is not. An agent is an application: it needs a step budget, error handling for tools that fail or return junk, a persistence story for long runs, permission scoping for anything that writes, and evaluation that measures whole trajectories rather than single responses. Teams that treat the loop as a prompting problem tend to ship a demo that works and a system that cannot be operated.