What is an Agent Loop?

This is the mechanical heart of what makes something an "agent" rather than a single-shot model call. A plain chatbot takes one input and produces one output; an agent instead runs a loop — often described as "reason, act, observe" — where the model's own output can be a request to call a tool, the system executes that tool, and the result is fed back into the model's context for the next iteration. The loop continues until the model produces a final answer or a stopping condition is hit. The practical challenges in an agent loop are knowing when to stop (an unbounded loop can burn cost and time without making progress), managing how much history accumulates in context as the loop runs long, and handling tool failures gracefully so one bad API response doesn't derail the whole task. Well-built loops include explicit budgets — a maximum number of steps or a time bound — and clear termination logic, distinct from the runtime timeouts that should never be applied to the underlying model calls themselves.