Agents & Tool Calling interview questions

An agent runs the ReAct loop: a Thought (reason about what to do next), an Action (call a tool), and an Observation (the result fed back). It keeps cycling — think, act, observe, think again — building up what it has learned, and it stops when it has gathered enough to write a final answer (or when a safety cap like a step limit forces it to). That loop is why an agent can answer things it couldn't possibly know up front: it discovers the pieces one at a time.

The model emits a request — the tool's name plus filled-in arguments as JSON — and then pauses; it never executes anything itself. Your code sees that request, runs the real function, and appends two things to the conversation: the model's own request and a tool_result tied to it by an id. Then you call the model again with the longer message history, and now it can read the result and decide the next move. The growing transcript is the agent's working memory.

Thought is the model reasoning out loud about what to do next — the "reason" half of ReAct. Action is the "act" half: an actual tool call (a function name plus arguments). Observation is what your code hands back — the tool's result. The model reads the observation, forms the next thought, and the cycle continues. Interleaving reasoning with actions (the 2022 ReAct paper's finding) beats doing either alone.