什么是图工程及其走红原因解析

Alex Martin 2026-07-21 9.5 KB 𝕏 文章列表

文章解释了从“循环工程”到“图工程”的技术演进。循环是简单的单一代理执行模式，而图（由节点、边和状态组成）通过可视化的流程图处理复杂逻辑和多代理协作。文章介绍了如何使用 LangGraph 构建第一个图，并指出在逻辑变得复杂时应从循环升级到图。

# Wtf is graph engineering and why is it going viral? **作者**: Alex Martin **日期**: 2026-07-20T15:30:57.000Z **来源**: [https://x.com/LoopOnChain/status/2079227569598300273](https://x.com/LoopOnChain/status/2079227569598300273) --- ![Image](/doc/3232/image/img_d8e4777f90c3.jpg) By the end of this you'll actually understand what a "graph" is, why everyone on your timeline suddenly won't shut up about them, and you'll have the code to build your first one, so stick around! Ok so here's what happened. Back in June, a guy named Peter Steinberger (@steipete, built OpenClaw, now at OpenAI, basically AI-agent famous) posted two sentences that did 8.4 million views: > "Here's your monthly reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." And everyone lost it. Addy Osmani over at Google wrote a whole essay calling it "Loop Engineering" the same day. Boris Cherny, the guy who literally built Claude Code, said basically "I don't prompt Claude anymore, I have loops running, they're the ones prompting Claude." So for a month, LOOPS were the thing. Everyone was building loops. Then last night Peter posts this: > "Are we still talking loops or did we shift to graphs yet?" And now everyone's quietly panicking about graphs. Here's the honest part: that tweet is half a joke. He's poking fun at how fast our little corner of the internet sprints from one buzzword to the next. But he's also not wrong. Graphs really are the next step up. And almost nobody is explaining them in plain english, so you get this low background dread that you're already behind. You're not. This one is genuinely simple. Let me show you. ## First, what a loop even is. A loop is the simplest possible agent. It goes: think, do something, look at what happened, think again. Over and over until the job's done. Reason, act, observe, repeat. You've felt this. It's Claude Code chugging away, running a command, reading the output, running the next one, until your feature is built. One agent, going in circles until it's finished. Loops are great. But they hit a ceiling. Because real work isn't one clean circle. Sometimes you need "if the tests pass, ship it, but if they fail, go fix them and come back." Sometimes you want three agents working at once and one boss agent pulling it together. Sometimes you need to stop and ask a human before doing something scary. A plain loop can't do that stuff cleanly. It all ends up buried in tangled if-statements inside your code that you can't actually see. That's where a graph comes in. ## What a graph actually is (the whole thing, in one picture) Draw a flowchart. Boxes connected by arrows. That's a graph. Seriously, that's the whole concept. Except this flowchart runs. The boxes do real work, and the arrows decide what happens next. There are exactly 3 pieces. Get these 3 and you get graphs, forever: 1. Nodes (the boxes). A node is one unit of work. "Call the LLM." "Run this tool." "Search the web." Each box does its one job and nothing else. 2. Edges (the arrows). An arrow points from one box to the next and decides where the work goes. And here's the good part: an arrow can point BACKWARD. Box B can hand the work back to Box A. That's a loop. A loop is literally just a graph with an arrow that curls back on itself. Arrows can also split. "If this, go here. If that, go there." That's a conditional edge, and it's the agent making a decision about its own next move. That split is the thing loops couldn't do cleanly. 3. State (the clipboard). There's one shared clipboard every box reads from and writes to. The conversation so far, which tools got called, whatever matters. Each box picks up the clipboard, does its thing, writes down what it learned, passes it on. That's it. Boxes, arrows, clipboard. Nodes, edges, state. Ok Alex, so what's actually different from a loop? A loop IS a graph. The simplest one. One box that keeps handing the clipboard back to itself. "Graph engineering" is just the name for what you do when your agent gets complicated enough that one box isn't enough, so you start drawing the real boxes and arrows instead of cramming everything into a while-loop you can't read. Branching. Multiple agents. A human checkpoint in the middle. All of it visible, all of it something you can point at when it breaks. That's the whole shift. Loops said "stop being the thing that prompts." Graphs say "draw the map before the whole thing turns to spaghetti." ## How to build your first graph (today, for real) The tool everyone uses is called LangGraph. It's free, it's got over 37,000 stars on GitHub, and your first graph is well under 50 lines of python. The classic first graph is the simplest useful agent there is: it thinks, and if it needs a tool it grabs one, then thinks again, until it's done. The famous "ReAct" pattern, except now you can see it as a graph instead of a black box. Here's the shape: • State: just the running list of messages. • Node 1, "reason": call the model. • The decision (conditional edge): did the model ask to use a tool? If yes, go to the tools node. If no, we're done. END. • Node 2, "tools": run whatever tool it asked for, then hand the clipboard back to "reason." That backward arrow from tools to reason? That's your loop. Living inside your graph. You just drew it instead of hiding it. The full copy-paste code is at the bottom. It runs. You'll have a working agent-graph in about ten minutes. ## The one honest catch. You don't need a graph for everything. If your agent is genuinely just one loop doing one thing until it's done, a graph is overkill and you should just write the loop. Don't graph-engineer your grocery list because Twitter told you to. You reach for a graph the moment you catch yourself writing "ok but if THIS happens do THAT instead, unless the other thing, in which case go back and..." That tangle in your head is a graph asking to be drawn. So here's what I'd actually do: • Play with a plain loop first. Really get it. (If loops don't click yet, that's your homework before this.) • Then build the little graph below. Change one node. Add one arrow. Watch it route. • The moment a loop starts getting tangled, that's your signal to graduate. Loops taught us to stop being the thing in the middle. Graphs teach us to draw the map before the map draws itself. Go build one. ## Here's the version to steal ``` Help me build my first graph, by teaching me to become a graph engineer with this first project: # Your first graph: a ReAct agent. # It thinks, uses a tool if it needs one, thinks again, until it's done. # pip install langgraph langchain-anthropic from typing import Annotated, TypedDict from langgraph.graph import StateGraph, START, END from langgraph.graph.message import add_messages from langgraph.prebuilt import ToolNode, tools_condition from langchain_anthropic import ChatAnthropic from langchain_core.tools import tool # 1. STATE: the shared clipboard. Here it's just the running list of messages. class State(TypedDict): messages: Annotated[list, add_messages] # A tool the agent is allowed to reach for. @tool def multiply(a: int, b: int) -> int: """Multiply two numbers.""" return a * b tools = [multiply] model = ChatAnthropic(model="claude-opus-4-8").bind_tools(tools) # 2a. NODE "reason": call the model with everything so far. def reason(state: State): return {"messages": [model.invoke(state["messages"])]} # Build the graph. graph = StateGraph(State) graph.add_node("reason", reason) graph.add_node("tools", ToolNode(tools)) # 2b. NODE "tools": runs whatever tool was asked for # 3. EDGES (the arrows): graph.add_edge(START, "reason") # start by reasoning graph.add_conditional_edges("reason", tools_condition) # tool asked? -> "tools", else -> END graph.add_edge("tools", "reason") # the backward arrow = your loop app = graph.compile() # Run it. for step in app.stream({"messages": [("user", "What's 12 times 7? Use your tool.")]}): print(step) ``` And if you don't want to read code, here's the prompt. Paste it into Claude and it'll build this WITH you and explain every line: ``` Build me my first LangGraph agent from scratch in Python. Use the three primitives: a State that holds the message list, a 'reason' node that calls the model, a conditional edge that routes to a 'tools' node if the model asked for a tool (otherwise END), and a 'tools' node that runs the tool and loops back to reason. Keep it under 50 lines, put a one-line plain-english comment on every node and edge so I understand each piece, then show me how I'd add a second tool and a human-approval checkpoint before a tool runs. ``` Follow me for more of this. I take the AI stuff everyone's panicking about and turn it into things you can actually build, minus the hype. If this made graphs finally click, hit like so it reaches the next person who's quietly confused. And send it to the one friend who keeps saying "graphs" in meetings and clearly doesn't know what they are. You know who. The best way to support me is with a comment below! ❤️ ## 相关链接 - [Alex Martin](https://x.com/LoopOnChain) - [@LoopOnChain](https://x.com/LoopOnChain) - [1.7K](https://x.com/LoopOnChain/status/2079227569598300273/analytics) - [@steipete](https://x.com/@steipete) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [11:30 PM · Jul 20, 2026](https://x.com/LoopOnChain/status/2079227569598300273) - [1,757 Views](https://x.com/LoopOnChain/status/2079227569598300273/analytics) --- *导出时间: 2026/7/21 14:04:46* --- ## 中文翻译 # 图工程到底是个啥，为什么突然爆火了？ **作者**: Alex Martin **日期**: 2026-07-20T15:30:57.000Z **来源**: [https://x.com/LoopOnChain/status/2079227569598300273](https://x.com/LoopOnChain/status/2079227569598300273) --- ![Image](/doc/3232/image/img_d8e4777f90c3.jpg) 读完这篇文章，你将真正理解什么是“图”，为什么你时间线上的每个人突然都在喋喋不休地谈论它，而且你将获得构建你的第一个图的代码。所以，请继续看下去！好吧，事情是这样的。今年六月，一个叫 Peter Steinberger 的人（@steipete，构建了 OpenClaw，现在在 OpenAI，基本上是 AI 代理圈的名人）发了两句话，获得了 840 万次浏览： > “这不仅是你的每月提醒：你不再应该去提示编码代理了。你应该设计循环来提示你的代理。” 所有人都炸锅了。Google 的 Addy Osmani 当天就写了一整篇文章，称之为“循环工程”。Boris Cherny，那个亲手构建了 Claude Code 的人，基本上是说“我不再提示 Claude 了，我有循环在运行，是它们在提示 Claude。” 所以整整一个月，循环就是一切。每个人都在构建循环。然后昨晚 Peter 发了这条推文： > “我们还在谈论循环，还是已经转向图了？” 现在所有人都在为图暗自恐慌。这里说实话：那条推文半是玩笑。他在取笑我们互联网的小圈子是如何从一个流行词冲刺到下一个流行词的。但他也没错。图确实就是进阶的下一步。几乎没有人用通俗的英语解释它们，所以你会有这种低微的背景焦虑，觉得自己已经落后了。你并没有。这东西其实很简单。让我展示给你看。 ## 首先，循环到底是什么。循环是可能存在的最简单的代理。它的流程是：思考，做点什么，观察发生了什么，再思考。如此往复，直到工作完成。推理，行动，观察，重复。你体验过这种感觉。就是 Claude Code 在轰隆隆地运行，执行一个命令，读取输出，运行下一个，直到你的功能构建完成。一个代理，一直转圈直到结束。循环很棒。但它们碰到了天花板。因为现实的工作并不是一个整洁的圆圈。有时你需要“如果测试通过，就发布，但如果失败，就去修复并回来。”有时你想让三个代理同时工作，一个主管代理把它们整合起来。有时你在做危险的事情之前需要停下来问一下人类。一个单纯的循环没法干净利落地做这些事。所有东西最后都埋藏在你代码里错综复杂的 if 语句中，你根本看不见。这就是图登场的时候了。 ## 图到底是什么（全貌，一张图搞定）画个流程图。用箭头连接的方框。那就是图。说真的，这就是全部概念。除了这个流程图会运行。方框做真正的工作，箭头决定下一步发生什么。确切地只有 3 个部分。搞定这 3 个，你就永远懂图了： 1. 节点（方框）。节点是一个工作单元。“调用 LLM。”“运行这个工具。”“搜索网络。”每个方框做它唯一的工作，不做别的。 2. 边（箭头）。箭头从一个方框指向下一个，并决定工作去哪。这是好消息：箭头可以指向**后方**。方框 B 可以把工作交回给方框 A。这就是循环。循环实际上就是一个箭头回指自己的图。箭头也可以分叉。“如果是这个，去这儿。如果是那个，去那儿。”这就是条件边，是代理在决定它自己的下一步动作。那个分叉正是循环没法干净利落地做到的事情。 3. 状态（剪贴板）。有一个共享的剪贴板，每个方框都从中读取和写入。目前为止的对话，调用了哪些工具，任何重要的事情。每个方框拿起剪贴板，做它的事，写下它学到的东西，传递下去。就是这样。方框，箭头，剪贴板。节点，边，状态。好吧 Alex，所以这跟循环到底有啥不同？循环**就是**一个图。是最简单的那种。一个不停地把剪贴板交回给自己的方框。 “图工程”只是个名字，指的是当你的代理变得复杂到一个方框不够用时，你开始画真正的方框和箭头，而不是把所有东西都塞进一个你读不懂的 while 循环里。分支。多个代理。中间有一个人类检查点。所有东西都是可见的，当它坏掉时，所有东西都是你可以指指点点的东西。这就是整个转变。循环说“别做那个负责提示的人。”图说“在整团乱麻变成意大利面之前，先画出地图。” ## 如何构建你的第一个图（今天，认真的）大家都在用的工具叫 LangGraph。它是免费的，在 GitHub 上有超过 37,000 颗星，而且你的第一个图绝对不超过 50 行 Python 代码。经典的第一个图是可能存在的最简单有用的代理：它思考，如果需要工具就抓一个，然后再思考，直到完成。著名的“ReAct”模式，除了现在你可以把它看作一个图而不是一个黑盒。它的结构是这样的： • 状态：只是运行中的消息列表。 • 节点 1，“reason”（推理）：调用模型。 • 决策（条件边）：模型要求使用工具了吗？如果是，去工具节点。如果否，我们完成了。结束。 • 节点 2，“tools”（工具）：运行它要求的任何工具，然后把剪贴板交回给“reason”。那个从工具指回推理的向后的箭头？那就是你的循环。生活在你的图里面。你只是画出了它，而不是把它藏了起来。完整的复制粘贴代码在底部。它能跑。大约十分钟后你就有了一个工作的代理-图。 ## 唯一的实诚陷阱。你不需要对所有事情都用图。如果你的代理真的只是一个循环做一件事直到完成，图就是杀鸡用牛刀，你应该直接写循环。别因为 Twitter 告诉你，你就对你的购物清单进行图工程。当你发现自己正在写“好吧但如果**这**发生就做**那**，除非是另一回事，那种情况下就回去……”的时候，就是你该伸手拿图的时候。你脑子里的那个乱团就是一个图请求被画出来。所以我实际上会这样做： • 先玩玩单纯的循环。真正搞懂它。（如果你还没理解循环，那是你做这个之前的家庭作业。） • 然后构建下面的小图。改变一个节点。加一个箭头。看它如何路由。 • 一旦循环开始变得纠缠不清，那就是你毕业的信号。循环教导我们别再做中间那个人。图教导我们在图把自己画出来之前先画出地图。去构建一个吧。 ## 这里是可供偷跑的版本 ``` 帮我构建我的第一个图，教我通过第一个项目成为一名图工程师： # 你的第一个图：一个 ReAct 代理。 # 它思考，如果需要工具就用一个，再思考，直到完成。 # pip install langgraph langchain-anthropic from typing import Annotated, TypedDict from langgraph.graph import StateGraph, START, END from langgraph.graph.message import add_messages from langgraph.prebuilt import ToolNode, tools_condition from langchain_anthropic import ChatAnthropic from langchain_core.tools import tool # 1. STATE：共享剪贴板。这里它只是运行中的消息列表。 class State(TypedDict): messages: Annotated[list, add_messages] # 代理被允许抓取的一个工具。 @tool def multiply(a: int, b: int) -> int: """两个数相乘。""" return a * b tools = [multiply] model = ChatAnthropic(model="claude-opus-4-8").bind_tools(tools) # 2a. 节点 "reason"：用目前所有的信息调用模型。 def reason(state: State): return {"messages": [model.invoke(state["messages"])]} # 构建图。 graph = StateGraph(State) graph.add_node("reason", reason) graph.add_node("tools", ToolNode(tools)) # 2b. 节点 "tools"：运行任何被请求的工具 # 3. 边（箭头）： graph.add_edge(START, "reason") # 从推理开始 graph.add_conditional_edges("reason", tools_condition) # 要求工具？-> "tools"，否则 -> END graph.add_edge("tools", "reason") # 向后的箭头 = 你的循环 app = graph.compile() # 运行它。 for step in app.stream({"messages": [("user", "12 乘以 7 是多少？用你的工具。")]}): print(step) ``` 如果你不想读代码，这里是提示词。把它粘贴到 Claude，它会和你**一起**构建这个并解释每一行： ``` 用 Python 从零开始帮我构建我的第一个 LangGraph 代理。使用三个原语：一个保存消息列表的 State，一个调用模型的 'reason' 节点，一个条件边（如果模型要求使用工具则路由到 'tools' 节点，否则 END），以及一个运行工具并循环回 reason 的 'tools' 节点。保持在 50 行以内，在每个节点和边上放一行通俗英语注释以便我理解每一部分，然后给我展示我要如何添加第二个工具以及在工具运行前增加一个人类批准检查点。 ``` 关注我获取更多这类内容。我把大家都在恐慌的 AI 东西变成你真正可以构建的东西，去掉那些炒作。如果这让图终于让你豁然开朗，点个赞，这样它就能触达下一个默默困惑的人。把它发给你那个在会议上不停说“图”但显然不知道那是啥的朋友。你知道我说的是谁。支持我的最好方式是在下面评论！❤️ ## 相关链接 - [Alex Martin](https://x.com/LoopOnChain) - [@LoopOnChain](https://x.com/LoopOnChain) - [1.7K](https://x.com/LoopOnChain/status/2079227569598300273/analytics) - [@steipete](https://x.com/@steipete) - [升级到 Premium](https://x.com/i/premium_sign_up) - [11:30 PM · Jul 20, 2026](https://x.com/LoopOnChain/status/2079227569598300273) - [1,757 Views](https://x.com/LoopOnChain/status/2079227569598300273/analytics) --- *导出时间: 2026/7/21 14:04:46*

相关文档