智能体运行时Agent runtime

心智模型Mental model

Agent 是「模型 + 工具 + 上下文 + 状态」的执行器；ReAct 循环把推理、工具调用和观察串成一个可中断、可流式、可持久化的会话。An agent is a model-plus-tools-plus-context executor; the ReAct loop turns reasoning, tool calls and observations into an interruptible, streamable, persistent conversation.

先学什么What to learn first

• 先掌握 `Agent.reply` / `reply_stream` / `observe` 的主流程，再看中间件的拦截点。Start with `reply`, `reply_stream` and `observe`, then follow middleware interception points.
• `RuntimeContext` 决定本次调用的用户、会话、权限和追踪信息。`RuntimeContext` carries user, session, permission and tracing information for one call.
• 状态通过 `AgentState` 与状态存储在外部分离，进程重启不丢会话。Conversation state lives outside the agent in `AgentState` stores, so restarts do not lose it.

生产常见坑Production pitfalls

• 把业务状态放在 Agent 实例字段里，水平扩容后状态不一致。Keeping business state in agent fields breaks horizontal scaling.
• 在工具里阻塞等待流式回复，导致事件循环饥饿。Blocking inside a tool while streaming starves the event loop.

进入对应闯关单元Practice this domain →