Thought, Action, and Observation

A standard LLM call is single-shot: input in, output out. An autonomous agent, however, runs in an iterative loop known as ReAct (Reasoning + Acting). At every turn, the agent thinks about what it needs, selects an action (such as executing a tool or querying a database), and waits for the environment's observation before deciding the next step.

The Three Phases of ReAct

• Thought: The model explicitly articulates its current state and reasoning (e.g. 'I need the user's order history before calculating their discount tier').
• Action: The model outputs a structured tool invocation with exact arguments (e.g. `getUserOrders({ userId: 'u_102' })`).
• Observation: Your runtime executes the tool and injects the actual result back into the prompt history as system/tool feedback.

type AgentStep =
  | { type: 'thought'; reasoning: string }
  | { type: 'action'; tool: string; args: Record<string, unknown> }
  | { type: 'observation'; output: unknown }
  | { type: 'finish'; response: string };