Agent loop
The base pattern behind every AI agent: LLM thinks → picks tool → runs → reads result → thinks again → until goal reached or max steps hit.
Original paper: ReAct (Yao et al., 2022) — the "Reasoning + Acting" loop. Thinking step + action step interleaved. Modern agent frameworks implement this (CrewAI, AutoGen, LangGraph, smolagents). Risks: infinite loops (max-step guards), expensive (many LLM calls), unreliable on complex tasks (>10 steps). OpenAI's o-series internalizes the loop in the model itself — no external orchestration.