What Is an Agent, Anyway?

A chatbot tells a user "I've scheduled that meeting with Sarah for next Tuesday at 2pm," and nothing actually happens. No calendar event exists anywhere. The model produced a plausible, confident sentence describing an action, because that is what a language model does, it generates plausible next tokens, and it has no built-in way to actually reach outside itself and act on the world. An agent is the architecture that closes exactly that gap.

Everything covered so far in this tier has been a single request in, single response out. An agent breaks that pattern in one specific way: instead of only being able to produce text, the model is given a set of tools it can choose to call (a calendar API, a search function, a database query), and instead of stopping after one response, the system runs a loop, feeding a tool's result back to the model so it can decide the next step, repeating until the task is genuinely complete.

The mechanism underneath this is function calling: alongside the prompt, you describe each available tool to the model as a schema (a name, a description of what it does, and the parameters it accepts). The model does not execute anything itself, it can never directly reach outside its own context. Instead, when it decides a tool is needed, it outputs a structured request naming the function and the arguments to call it with. Your application code is what actually executes that call and returns the real result back into the conversation.

Real tasks rarely resolve in a single tool call. Scheduling that meeting properly means checking availability first, and only booking once that check comes back clear, two full reason-act-observe cycles before the agent has enough grounded information to give a real answer. Step through the trace below and notice the shape repeating: think about what is needed, act by calling a tool, observe the real result, then think again with that new information in hand.

This is still, relatively speaking, the simplest kind of agent: a fixed, predictable sequence of two tool calls. The full Agents & Tool Use tier ahead builds this out considerably further: the ReAct pattern named and formalized (explicitly interleaving reasoning about what to do next with the actual acting, rather than jumping straight to a tool call), agents that plan multi-step tasks and decompose them, agents that maintain memory across an extended task, multiple agents collaborating, and the hardening work (error handling, evaluation, orchestration) that turns a working demo into something safe to run in production.