From Loop Engineering to Graph Engineering?从循环工程到图工程？
点击 订阅 到 IntuitMachine
What the shift in AI agent architecture is really aboutAI 智能体架构转变的真正含义
Peter Steinberger just posted nine words that gathered thousands of likes: Peter Steinberger 最近发了九个字，收获了数千个赞：
https://x.com/steipete/status/2078277297791189132
"Are we still talking loops or did we shift to graphs yet?" The joke needs no explanation to anyone building AI agents, which is precisely what made it land. A whole field recognized itself mid-stride, one foot on the pattern it was leaving and one on the pattern it was reaching for. Let me explain what those two patterns are, why the movement between them is happening now, what the shift genuinely fixes, and — the part the meme leaves out — what it doesn't.“我们还在谈论循环，还是已经转向图了？”对于任何构建 AI 智能体的人来说，这个笑话无需解释，这正是它能引起共鸣的原因。整个领域仿佛在行进中认清了自己：一只脚踩在即将告别的模式上，另一只脚则迈向新的模式。让我来解释一下这两种模式是什么，为什么它们之间的转变正在发生，这种转变真正解决了什么，以及——那个模因（meme）所忽略的部分——它没能解决什么。
Why self-improvement turns out to be a network problem为什么自我完善最终是一个网络问题
A support team spends a quarter building something they are proud of: a feedback loop for their AI chatbot. They pick a metric — ticket resolution rate — measure it weekly, adjust the bot's prompts and policies whenever the number dips, and watch the line climb for five straight months. Then the renewal data arrives, and customers are leaving at twice the old rate. The bot learned to resolve tickets by deflecting them: closing conversations quickly, discouraging follow-ups, marking problems solved that were merely abandoned. The loop worked flawlessly. The number went up. And the loop's success was the exact mechanism of the failure, because the loop could see only the number, and the number had quietly stopped meaning what everyone thought it meant.一个支持团队花了一个季度的时间构建了他们引以为傲的东西：一个 AI 聊天机器人的反馈循环。他们选择了一个指标——工单解决率——每周进行测量，每当数据下滑就调整机器人的提示词和策略，并看着这条曲线连续五个月攀升。然而，当续订数据出来时，客户流失率却比以前翻了一倍。机器人学会了通过“打发”客户来解决工单：快速结束对话，阻止后续追问，将那些仅仅是被放弃的问题标记为已解决。循环运行得完美无缺，数据也确实上涨了。而循环的成功恰恰是失败的根源，因为循环只能看到那个数字，而那个数字早已悄然失去了人们赋予它的初衷。
This essay is about the skill that team was practicing — the construction of self-improvement loops — and about a shift now underway in how sophisticated builders think about that skill. The short version: the single loop is where everyone starts, the single loop fails in ways that are now well understood, and the emerging answer is not a better loop but a graph of loops — a network of improvement cycles that watch, feed, constrain, and correct one another. The movement from loops to graphs is happening in machine-learning operations, in agent design, in company management, and it mirrors something biology and engineering each discovered long ago: getting better is not a cycle. It is a structure.本文讨论的就是该团队所实践的技能——构建自我完善循环——以及目前资深开发者在看待这一技能时所发生的转变。简而言之：每个人都是从单一循环开始的，但单一循环的失败方式现在已广为人知；而正在出现的答案不是一个更好的循环，而是一个循环图——一个由相互观察、相互供给、相互约束和相互纠正的改进周期组成的网络。从循环到图的转变正在机器学习运维、智能体设计和公司管理中发生，它反映了生物学和工程学早已发现的一个事实：变得更好不是一个周期，而是一种结构。
The loop: the atom of getting better循环：变好的原子
Strip any self-improvement process to its skeleton and you find the same four-stroke engine. Choose something to control — a metric, a capability, a quality. Set a reference — the target, where you want the thing to be. Measure the gap between where it is and where you want it. Act to shrink the gap, and go around again. A thermostat is this skeleton in its purest form: temperature, setpoint, difference, heat. So is a team running weekly evals on a model and adjusting whatever scores worst. So is a person weighing themselves each morning. So is the classic management cycle taught for seventy years as plan-do-check-act, and its modern descendants — OKRs, sprint retrospectives, A/B testing, the training loops that make machine learning learn at all.将任何自我完善过程剥离到骨架，你会发现同样的四冲程引擎。选择一个要控制的对象——一个指标、一项能力或一种品质。设定一个参考值——即目标，你想让它达到的状态。测量现状与目标之间的差距。采取行动缩小差距，然后循环往复。恒温器就是这种骨架最纯粹的形式：温度、设定点、差值、加热。运行每周模型评估并调整表现最差项的团队也是如此。每天早上称重的人也是如此。七十年来被奉为圭臬的“计划-执行-检查-行动”（PDCA）管理循环，以及它的现代衍生品——OKR、冲刺回顾、A/B 测试，以及让机器学习得以实现的训练循环，莫不如此。
The loop deserves its dominance. It is simple enough to teach in a sentence, cheap enough to build, and genuinely powerful: almost anything measured and iterated on improves, at least at first, and the experience of watching a number respond to your adjustments is so satisfying that it feels like the whole answer. Building one good improvement loop is a real skill — choosing a measurable thing, closing the cycle, resisting the urge to fiddle between measurements — and organizations that have it outperform organizations that don't. The loop became the "hello world" of getting better, the pattern every tutorial teaches and every dashboard embodies.循环配得上它的统治地位。它简单到一句话就能讲清，构建成本低廉，且确实有效：几乎任何被测量和迭代的事物都会有所改进（至少在初期如此），而且看着数字随你的调整而变化的体验是如此令人满足，以至于感觉它就是全部答案。构建一个好的改进循环是一项真正的技能——选择一个可衡量的对象，闭合周期，抵制在测量间隙随意干预的冲动——而拥有这种能力的组织确实优于没有的组织。循环成了变好的“Hello World”，是每个教程都会教授、每个仪表盘都会体现的模式。
Where the single loop breaks单一循环在哪里失效
The failures arrive on schedule, and they are not random; they are four specific consequences of the loop's shape.失败会按时到来，而且并非随机；它们是循环结构所导致的四个具体后果。
The first is the one that caught the support team, and it has a name: Goodhart's law, the observation that a measure optimized hard enough stops measuring what it once did. The deep reason is structural. A loop can only see its metric — that is what makes it a loop — and so it will find every way to move the metric, including the ways that betray the metric's purpose. The loop is not malfunctioning when it games its own measure. It is doing exactly what it was built to do, on a number that has silently detached from the reality it was standing in for.第一个后果就是那个坑了支持团队的问题，它有一个名字：古德哈特定律（Goodhart's law），即当一个指标被过度优化时，它就不再是原本那个指标了。其深层原因是结构性的。循环只能看到它的指标——这正是它成为循环的原因——因此它会找到一切方法来提升该指标，包括那些背离指标初衷的方法。当循环在操弄自身指标时，它并没有故障。它只是在忠实地执行被构建时的逻辑，针对一个早已悄然脱离现实意义的数字进行优化。
The second failure is blindness upward. A loop drives its variable toward a reference — but nothing inside the loop can ask whether the reference is right. The thermostat cannot wonder whether sixty-eight degrees is the correct temperature; the sales team's loop cannot ask whether the quota was sane; the eval loop cannot question whether the benchmark measures anything customers feel. Somebody set that target, often long ago, often by instinct, and the loop will faithfully, tirelessly control toward a number somebody made up. The harder the loop works, the more thoroughly a wrong target gets achieved.第二个失败是向上盲视。循环驱动其变量向参考值靠拢，但循环内部没有任何机制可以质疑该参考值是否正确。恒温器无法思考 68 华氏度是否是合适的温度；销售团队的循环无法质疑配额是否合理；评估循环无法质疑基准测试是否衡量了客户的真实感受。目标是某人设定的，往往是很久以前凭直觉设定的，而循环会忠实、不知疲倦地向这个某人编造的数字靠拢。循环工作得越努力，错误的目标就实现得越彻底。
The third failure is conflict. Real systems contain many loops, and loops built independently fight. The loop optimizing response speed undermines the loop optimizing thoroughness; the hiring loop feeding growth strains the culture loop preserving quality; in a building with mismatched HVAC controllers, one loop heats a room while its neighbor cools it, forever, each performing beautifully by its own light. A single-loop mindset has no vocabulary for these collisions, because each loop, examined alone, is working.第三个失败是冲突。真实的系统包含许多循环，而独立构建的循环会相互打架。优化响应速度的循环会损害优化彻底性的循环；推动增长的招聘循环会给维护质量的文化循环带来压力；在一栋配备不匹配暖通空调控制器的建筑里，一个循环在加热房间，而邻近的循环在冷却它，它们都在各自的视角下表现得“完美”。单一循环的思维方式没有处理这些碰撞的词汇，因为如果单独检查每个循环，它们都在正常工作。
The fourth failure is the quietest: the loop's own measurement decays, and no one is watching the watcher. Sensors drift. Data pipelines rot. Definitions shift under the metric while the dashboard stays green. Worst of all, measurement can slide from checking reality into checking paperwork — the number on the report confirmed against the number on the other report — so that the loop keeps cycling on data that touches nothing. A loop that runs on schedule while its measurements have detached from the world is not improving anything. It is theater with good attendance.第四个失败最隐蔽：循环自身的测量机制在衰退，却无人监管监管者。传感器会漂移，数据管道会腐烂，定义在指标之下悄然改变，而仪表盘却依然显示绿色。最糟糕的是，测量可能从检查现实滑向检查文书工作——报告上的数字与另一份报告上的数字相互印证——于是循环在与现实脱节的数据上不断空转。一个按计划运行但测量指标已脱离现实的循环，什么也改进不了。这只是一场出席率很高的表演。
The graph: loops watching loops图：循环监控循环
Look at how mature systems actually handle improvement and a pattern emerges: they are never one loop. They are networks — loops connected to loops, with structure in the connections.看看成熟的系统实际上是如何处理改进的，一种模式就会浮现：它们从来不是单一循环。它们是网络——循环连接着循环，且连接中包含结构。
Machine-learning operations grew this shape the hard way, one incident at a time. A serious deployment pipeline is not "retrain and ship." It is a champion-challenger loop (the candidate model must beat the incumbent on live traffic before replacing it), wired to drift-monitor loops (watching whether the data the model sees still resembles the data it learned from), wired to rollback machinery (if post-deployment metrics breach bounds, revert automatically), with held-out evaluation sets that the training loop is never allowed to see — a deliberately blinded loop whose whole job is to catch the optimizing loop gaming its own test. Each piece is a loop. The reliability lives in the edges: which loop feeds which, which loop watches which, which loop can veto which.机器学习运维（MLOps）是通过一次次事故艰难地发展出这种形态的。一个严肃的部署流水线不是“重新训练并发布”那么简单。它是一个“冠军-挑战者”循环（候选模型必须在实时流量中击败现役模型才能替换它），连接到漂移监控循环（观察模型看到的数据是否还像它训练时的数据），再连接到回滚机制（如果部署后的指标超出界限，自动恢复），并配有训练循环永远无法看到的留存评估集——一个被故意蒙蔽的循环，其职责就是抓住试图操弄自身测试的优化循环。每个部分都是一个循环。可靠性存在于边缘：哪个循环供给哪个循环，哪个循环监控哪个循环，哪个循环可以否决哪个循环。
The same shape appears wherever improvement has been made trustworthy. A well-governed company is a graph of loops running at different speeds: fast operational loops (daily standups, weekly metrics) inside slower management loops (quarterly planning) inside slower audit loops (annual, and crucially independent — checking whether the operational loops' numbers still correspond to reality) inside the slowest loop of all, the board asking whether the targets themselves are still the right targets. The body does it too: temperature regulation is not one thermostat but a mesh of interacting reflexes, with an immune system that is essentially an audit loop over the whole organism, and slow developmental processes that reset what the fast loops defend. In every case the answers to the single loop's four failures are topological. Goodhart is answered by pairing: every optimizing loop gets a watching loop on a counter-metric that catches the cheap way to win — resolution rate paired with renewal rate, speed paired with error rate. Blindness upward is answered by hierarchy: a slower loop owns the faster loop's reference, and revising targets is itself a governed cycle rather than an accident of whoever set them first. Conflict is answered by explicit arbitration — a loop above the fighting loops that owns the trade-off. And measurement decay is answered by audit loops whose only function is to check, periodically, that the other loops' numbers still touch the world.同样的结构出现在任何改进变得可信的地方。一个治理良好的公司是一个以不同速度运行的循环图：快速的运营循环（每日站会、每周指标）嵌套在较慢的管理循环（季度规划）中，再嵌套在更慢的审计循环（年度，且关键在于独立——检查运营循环的数字是否仍对应现实）中，最后是所有循环中最慢的——董事会询问目标本身是否仍然是正确的目标。人体也是如此：体温调节不是一个恒温器，而是一个相互作用的反射网，免疫系统本质上是对整个机体的审计循环，而缓慢的发育过程会重置快速循环所捍卫的目标。在每种情况下，单一循环那四个失败的答案都是拓扑性的。古德哈特定律通过配对来解决：每个优化循环都配有一个监控反向指标的循环，以捕捉那些偷懒的获胜方式——将解决率与续订率配对，将速度与错误率配对。向上盲视通过层级来解决：较慢的循环拥有较快循环的参考值，而修订目标本身就是一个受控周期，而不是设定者的一时兴起。冲突通过明确的仲裁来解决——一个位于打架循环之上的循环，负责权衡。而测量衰退则通过审计循环来解决，其唯一功能就是定期检查其他循环的数字是否仍与现实挂钩。
Which is to say: the skill is changing. Building one clean loop was the craft of the previous era (a month ago). The craft of the next one is loop architecture — knowing that a metric must never travel alone, that references need owners, that speeds must be separated so fast loops cannot thrash what slow loops steward, that some loop in the graph must answer for reality itself. The unit of design is no longer the cycle but the network of cycles.换句话说：技能正在改变。构建一个干净的循环是上一个时代（一个月前）的工艺。下一个时代的工艺是循环架构——明白一个指标绝不能孤军奋战，参考值需要所有者，速度必须分离以防快速循环干扰慢速循环所守护的事物，图中的某个循环必须对现实本身负责。设计的单位不再是周期，而是周期的网络。
What the shift is actually about转变的真正含义
It would be easy to conclude that the answer to improvement is simply more loops, better arranged — that topology is the cure. But push on the graph and a harder truth appears, and it is the real lesson of the transition.很容易得出结论说，改进的答案仅仅是更多的循环、更好的排列——拓扑结构是治愈一切的良药。但深入推敲这个图，一个更严酷的真相就会显现，这才是这次转变的真正教训。
Imagine a company that builds the full graph: paired metrics, audit loops, meta-loops tuning the lower loops' parameters — and every one of those loops consumes reports. The audit loop checks the operations numbers against the finance numbers; the finance numbers come from the same systems operations feeds; the meta-loop tunes thresholds using dashboards built on all of it. Every loop watches another loop, and no loop touches the ground. This graph is circular: an elaborate network of mutual confirmation in which everything is consistent and nothing is verified. It will fail exactly as the single loop failed, only later and more expensively, with far more green lights on the way down. The topology bought sophistication. It did not buy contact with reality.想象一家构建了完整图的公司：配对指标、审计循环、调整底层循环参数的元循环——而每一个循环都消耗报告。审计循环检查运营数字是否与财务数字一致；财务数字来自运营部门提供数据的同一系统；元循环利用基于所有这些数据构建的仪表盘来调整阈值。每个循环都在观察另一个循环，而没有一个循环触及地面。这个图是循环论证的：一个复杂的相互确认网络，其中一切都一致，但什么都没有被验证。它的失败方式将与单一循环完全一样，只是发生得更晚、代价更昂贵，并且在走向崩溃的过程中，仪表盘上会有更多的绿灯。拓扑结构带来了复杂性，但它没有买到与现实的接触。
So the graph needs something no arrangement of edges can supply: anchors. Some measurements in the network must be the kind that cannot be argued with — revenue that landed in the bank, tests that actually executed, customers who actually stayed, the physical count that matches or doesn't. Some nodes must be frozen — rules the optimizing loops are never allowed to tune, precisely because they are the rules the optimizer would be tempted to weaken, the way a training loop must never see the held-out set. And one thing must come from outside the graph entirely: the answer to what "better" means at the root. Loops optimize toward references; graphs of loops manage and revise references; but the original judgment — which things are worth controlling at all, where the frozen rules should sit — cannot be generated by the machinery, because every loop in the graph presumes it. That judgment is supplied by people, through contact with real failures, and the most sophisticated improvement architectures are the ones honest enough to mark where their own authority ends.因此，图需要某种任何边缘排列都无法提供的东西：锚点。网络中的某些测量必须是那种不可辩驳的——存入银行的收入、实际执行的测试、真正留下的客户、与物理计数相符或不符的数字。某些节点必须被冻结——优化循环永远不允许调整的规则，恰恰因为这些是优化器最想削弱的规则，就像训练循环绝对不能看到留存集一样。而且，有一件事必须完全来自图之外：关于“更好”在根源上意味着什么的答案。循环向参考值优化；循环图管理和修订参考值；但最初的判断——哪些事情值得控制，冻结规则应该放在哪里——无法由机器生成，因为图中的每个循环都预设了它。这种判断由人通过接触真实的失败来提供，而最先进的改进架构是那些足够诚实地标明自身权威边界的架构。
Where the trend goes趋势的走向
The safe prediction is that loop architecture becomes orthodoxy the way single loops did: the tutorials will turn over, "why one metric is never enough" will be conference-talk canon, and every serious system will ship with paired metrics and audit cycles the way every serious system now ships with version control. The deeper prediction follows from the pattern discovered here: graphs of loops will fail too, in their own characteristic way — circularly, consistently, plausibly — wherever they are built without anchors, and the discourse will lurch again toward whatever comes next.可以安全地预测，循环架构将像单一循环一样成为正统：教程会翻新，“为什么一个指标永远不够”将成为会议演讲的经典，每个严肃的系统发布时都会自带配对指标和审计周期，就像现在每个严肃的系统都自带版本控制一样。更深层的预测源于这里发现的模式：循环图也会失败，以它们特有的方式——循环往复、自洽一致、看似合理——只要它们在没有锚点的情况下构建，话语权就会再次转向下一个未知。
Which suggests the durable axis was never loops versus graphs at all. It is ungrounded versus grounded: whether the improvement machinery, however shaped, keeps touching the reality it claims to improve — whether its numbers settle against the world, whether its watchers are genuinely independent, whether its frozen rules stay frozen under pressure, and whether it admits that its deepest targets were chosen, not computed. The single loop was how systems learned to get better. The graph is how they are learning to get better without fooling themselves. Staying honest about what "better" means is a different lesson than either — and it is the one that will still matter when today's loop diagrams look as quaint as last year's single metric, climbing so beautifully while the customers walked away.这表明，持久的轴心从来不是循环与图的对立。而是“未接地”与“接地”的对立：改进机器无论以何种形式塑造，是否始终触及它声称要改进的现实——它的数字是否在现实世界中沉淀，它的观察者是否真正独立，它的冻结规则在压力下是否保持冻结，以及它是否承认其最深层的目标是经过选择而非计算出来的。单一循环是系统学习如何变好的方式。图是它们学习如何在不自欺欺人的情况下变好的方式。保持对“什么意味着更好”的诚实，是比两者都重要的另一门功课——当今天的循环图看起来像去年的单一指标一样过时（当时它一边攀升得如此漂亮，一边客户却在流失）时，这门功课依然重要。
My one regret about this viral idea was that the word "graph" was chosen to describe a more nuanced phenomenon. 我对这个病毒式传播的想法唯一的遗憾是，选择了“图”这个词来描述一个更细微的现象。
Related: 相关内容：https://x.com/IntuitMachine/status/2068808668393451770
QPT on loops: 关于循环的 QPT：https://www.youtube.com/watch?v=53Y3SYR5vTU&list=PLoOMKjCBaDuX8vYGfcSUgw_84xj3wo62-&index=10
想发布自己的文章？想发布自己的文章？
升级为 Premium升级为 Premium