ReAct Loop

ReAct (Reasoning + Acting) is the core operating mode of an agent: reasoning → action (tool call) → observation of the result → reasoning again. The loop repeats until the task is completed. The trajectory is a continuously accumulating history of messages (user / assistant / tool), and with each LLM call, it sees the entire trajectory along with a static prefix. Agent context = static prefix + trajectory.

Related: [Agent loop in code], [Context as the agent’s eyes]