四年，行业造了七个「X Engineering」：prompt → context → spec → harness → loop，到 fleet。看着像术语通胀，其实是一栋楼——每个新词优化的层，恰好在上一个词的上一层抽象。机制就是这套系列的内核，加上一根时间轴：模型每强一档，人手工操作的那层被产品吸收，人的杠杆点就上移一层。这一卷，就是内核第②步在四年里的轨迹；底层是控制论的复活。
In four years the field minted seven "X Engineering" terms: prompt → context → spec → harness → loop, and on to fleet. It looks like terminology inflation; it is actually one building. Each new term optimizes the layer exactly one abstraction above the last. The mechanism is this series' kernel plus a time axis: every time models gain a notch, the layer humans were hand-operating gets absorbed by products, and the human leverage point climbs a floor. This volume is the trace of the kernel's step ② over four years; underneath, it is cybernetics reborn.
把七个词按时间排开，看到的不是泡沫，而是一条单调上行的轨迹。每个新词的优化对象，都在前一个词的上一层抽象——没有一次跳跃是随机的。
Lay the seven words out by date and you see not a bubble but a monotonic upward trajectory. Each new term optimizes one abstraction above the last; no jump is random.
表面上这是术语通胀：每隔几个月，一位有影响力的工程师发条推文，一个新「X Engineering」便诞生，配套仓库、CLI 与嘲讽视频在七十二小时内集齐——2026 年 6 月的 loop engineering 把这周期压到不足一周。
On the surface it is terminology inflation: every few months an influential engineer posts, a new "X Engineering" is born, and the repo, CLI, and mocking video assemble within seventy-two hours. June 2026's loop engineering compressed that cycle to under a week.
但从单次调用的措辞，到上下文窗口里的信息分布，到意图的源文件，到单个智能体的运行环境，再到驱动智能体的闭环系统本身——模型能力每抬升一档，人原本手工操作的那层就被产品吸收，人被推向更高一层，新的那层需要一个名字。这一卷不追术语热度，它把七个词放回一栋楼里，判定哪些是承重结构、哪些只是装修。
But from the wording of a single call, to the distribution of information in the context window, to the source file of intent, to a single agent's runtime, to the closed loop that drives the agents: each time capability rises a notch, the layer humans hand-operated gets absorbed by products, humans are pushed up a level, and the new level needs a name. This volume does not chase the hype; it puts the seven words back into one building and judges which floors are load-bearing and which are just decoration.
系列内核：执行变充裕 → 判断退守 → 上下文成基设 → 人回归意义。这一卷给它加一根时间轴——退守的那个节点，会随能力提升沿抽象层级单向上移。这栋楼，就是第②步四年来的轨迹。
The series kernel: execution becomes abundant → judgment retreats → context becomes infrastructure → people return to meaning. This volume adds a time axis: the node it retreats to migrates monotonically up the abstraction stack as capability rises. The building is the trace of step ② over four years.
下面的图纸：SHEET 02 把七个词排成一栋楼，SHEET 03 指出它其实是控制论复活，SHEET 04 解剖当前顶层（Loop）的五件套，SHEET 05 区分它与自我迭代智能体，SHEET 06 用唯一判据筛出承重楼层，SHEET 07 给趋势、预测与实践。最后一件仪器，帮你定位自己在第几层。
The sheets ahead: SHEET 02 stacks the seven words into a building, SHEET 03 shows it is cybernetics reborn, SHEET 04 dissects the five pieces of the current top floor (Loop), SHEET 05 separates it from self-improving agents, SHEET 06 filters out the load-bearing floors by one criterion, and SHEET 07 gives trends, predictions, and practice. A final instrument locates which floor you are on.
左是楼层与它优化的对象，右是它在控制论里的对应组件。Agentic engineering 不是某一层，它是整栋楼的名字。这栋楼的电梯，只往上开。
Left is the floor and what it optimizes; right is its counterpart component in control theory. Agentic engineering is not a floor; it is the name of the whole building. Its elevator only goes up.
每上一层，人就把下一层交给产品：措辞被更强的指令遵循消化，上下文管理被自动压缩与检索消化，环境搭建正被开箱即用的 harness 消化。判断哪层承重、哪层装修，标准只有一条——它压住的约束，会不会随模型变强而消失。注意力有限（2F）、意图无人能替你定义（3F）、行动需要边界（4F），这三个约束是物理性的；围绕它们的学科不会过时，只会换名字。
Each floor up, humans hand the one below to products: wording is digested by stronger instruction-following, context management by automatic compression and retrieval, environment setup by off-the-shelf harnesses. The single test of load-bearing vs decoration: does the constraint it pins disappear as models get stronger? Finite attention (2F), intent that no one can define for you (3F), and action that needs boundaries (4F) are physical constraints; the disciplines around them do not expire, they only change names.
触发、执行、验证、状态持久化、生成者与检查者分离、监督者警觉性衰减——这些词在控制论、Sheridan 的监督控制理论与人因工程里，已存在五十年。行业正在重新发明反馈控制系统，只是没人这么叫。
Triggering, execution, verification, state persistence, generator/checker separation, supervisor vigilance decay: these have existed for fifty years in cybernetics, Sheridan's supervisory control theory, and human-factors literature. The field is reinventing feedback control systems, without calling it that.
看 SHEET 02 的右列——每一层都精确对应一个控制论组件：单次控制信号、传感器、设定值、执行器约束、闭环反馈、层级控制。这不是巧合。一个能跑的智能体循环，本质就是一套反馈控制系统：有设定值（spec）、有传感器（context）、有执行器（harness）、有闭环（loop）、有监督者（人）。
Look at the right column of SHEET 02: each floor maps precisely to a control-theory component – single control signal, sensor, setpoint, actuator constraint, closed-loop feedback, hierarchical control. This is no coincidence. A working agent loop is, at bottom, a feedback control system: a setpoint (spec), sensors (context), actuators (harness), a closed loop, and a supervisor (the human).
这给了我们一把尺子：凡是控制论里早有名字的东西，都会留下；凡是只为某次模型短板而临时补的技巧，都会被下一代模型消化。更要紧的是它点破一个被研究了四十年的陷阱（SHEET 05 展开）：系统越可靠，监督者越盯不住。
This hands us a ruler: whatever already has a name in cybernetics will stay; whatever is a stopgap for one generation's model gap will be digested by the next. More importantly, it names a trap studied for forty years (expanded in SHEET 05): the more reliable the system, the less the supervisor can watch it.
一个学科若只在控制论里换了层皮，它就会留下；若它只是某代模型的创可贴，它就会被吸收。名字会换，反馈控制的骨架不会。
A discipline that merely re-skins a control-theory idea will stay; one that is a band-aid for a single model generation gets absorbed. Names change; the feedback-control skeleton does not.
一个能跑的 loop（按 Osmani 的拆解）= 五个原语 + 一处外部记忆。Claude Code 与 Codex 已各自配齐、且一一同构——这意味着 loop 设计正成为工具无关的可迁移技能。
A working loop (per Osmani's breakdown) = five primitives + one external memory. Claude Code and Codex each have all five, isomorphically, which means loop design is becoming a tool-independent, portable skill.
把这六样配齐，你就有了一个能无人值守跑起来的循环。注意 05 那堵承重墙——没有独立验证，循环只是在用惊人的速度放大自己的错误。
Assemble these six and you have a loop that runs unattended. Note the load-bearing wall at 05: without independent verification, a loop merely amplifies its own mistakes at remarkable speed.
Loop 常被误认作自我改进智能体的同类。分界线只有一条：循环的结构由谁定义。四档共享同一副骨架——生成器 + 独立验证器 + 外部持久状态，即变异 / 选择 / 保留。
A loop is often mistaken for a self-improving agent. The line is single: who defines the loop's structure. All four levels share one skeleton – generator + independent verifier + external persistent state, i.e. variation / selection / retention.
编码 loop 用测试与 CI 做选择压力——廉价、确定、毫秒级；科研 loop 用实验与自然做选择压力——昂贵、有噪、以周计。这就是编码成为第一个被 loop 化的智力劳动的全部原因：它验证成本最低。哪个领域的验证成本降下来，哪个就是下一个。
A coding loop uses tests and CI as selection pressure: cheap, deterministic, millisecond-scale. A research loop uses experiments and nature: expensive, noisy, week-scale. That is the whole reason coding is the first intellectual labor to be loop-ified – its verification cost is the lowest. Whichever domain's verification cost drops next is the next to go.
从第Ⅱ档到第Ⅲ档，回应着 Bainbridge 1983《自动化的反讽》：系统越可靠，监督者警觉性衰减越快，而恰在最需要接管的异常时刻，人已丢失情境感知。「人在环上」注定盯不住。Loop 的回答不是要人盯得更紧，而是把验证写进结构（checker 子代理、可机检条件），再把人的介入从实时监督改成异步分诊——人不再需要与机器的时钟同步。
The jump from Ⅱ to Ⅲ answers Bainbridge's 1983 "Ironies of Automation": the more reliable the system, the faster supervisor vigilance decays, and exactly when takeover is needed, the human has lost situational awareness. "Human on the loop" is doomed to not watch. The loop's answer is not to watch harder but to write verification into the structure (checker subagents, machine-checkable conditions), and move human intervention from real-time supervision to asynchronous triage: the human no longer has to keep the machine's clock.
判据唯一：是否压住一个不随模型变强而消失的约束。压住的活下来，没压住的被吸收成素养。整张判定表都站在三个硬约束之上。
One criterion: does it pin a constraint that does not vanish as models improve? Those that do survive; those that do not get absorbed into literacy. The whole table stands on three hard constraints.
| 概念Concept | 押的约束Constraint | 判定Verdict |
|---|---|---|
| Prompt engineering | 模型对指令敏感（正消失）instruction-sensitivity (fading) | 吸收为素养absorbed as literacy |
| Context engineering | 注意力有限 · context rotfinite attention · context rot | 长期有效long-term |
| Spec engineering | 意图不可替代intent irreplaceable | 原理长存principle endures |
| Harness engineering | 行动需边界 · 知识需注入action needs bounds · knowledge injected | 长期有效long-term |
| Loop engineering | 验证昂贵 · 闭环反馈verification cost · closed loop | 长期有效long-term |
| Agentic engineering | 整个时代的总称name of the era | 伞形词留下umbrella stays |
| Vibe coding / engineering | 文化现象 / 败于命名之争culture note / lost the naming war | 时代注脚a footnote |
名字还会换，栈不会。三条趋势定方向，一条实践闭环给动作——核心一句：定位你在第几层，把下层交出去，把判断花在更上一层。
The names will keep changing; the stack will not. Three trends set direction, one practice loop gives the moves. The core: locate your floor, hand the floors below away, spend judgment one floor up.