1. The Human Is Not the User. The Human Is in the Loop.

人类不是用户。人类在回路中。

For most of AI's short history, the human was the user: the one who opened the product, typed the request, and received the output. The entire system was organized around the human's attention as the primary resource being served.

That is no longer the default shape.

In 2026, increasingly large portions of workflows run agent-to-agent, with tool calls, delegated subtasks, and automated decisions stacked inside automated decisions. The human still exists in this system — but not as the initiator of every action. The human is now an interrupt handler. A validator. An authority who must sign off on things that exceed the agent's permission boundary. A moral participant who cannot be removed from consequential decisions even when their involvement slows everything down.

This distinction matters operationally.

When the human was the user, your job was to serve their requests well. When the human is in the loop, your job is more delicate: you must invoke them precisely, protect their attention, preserve their sense of agency, and return control to the workflow without eroding the relationship that makes future invocations possible.

Most agent failures in production do not happen because the model was wrong. They happen because the model queried the human at the wrong moment, in the wrong frame, with the wrong amount of context — or did not query the human at all when it should have.

在 AI 短暂的发展史里，人类始终是那个用户——打开产品、输入请求、等待输出。整套系统的设计前提，是人的注意力是最核心的被服务对象。

这个前提已经不再成立。

2026 年，大量工作流跑在智能体之间：工具调用、子任务委派、自动决策套着自动决策。人类还在这个系统里，但不再是每个动作的起点。现在，人类是中断处理者，是校验员，是在智能体权限边界之外必须拍板的那个人，是哪怕拖慢流程也不能从高风险决策中移除的道德责任方。

这个区别，落到实操层面非常关键。

人类是用户的时候，你的工作是好好响应他的请求。人类处于环路之中，事情就复杂了：你得在对的时机打扰他，不能乱，不能频繁；要护住他的注意力；要让他觉得自己仍然掌控着什么；控制权交还工作流之后，还不能伤了这段协作关系——毕竟下次还要调用他。

生产环境里的智能体故障，大多数不是模型判断错了。问题出在：在错误的时机问了人，用了错误的框架，给的上下文太多或太少——或者本该问的时候，压根没问。