Anthropic发布Managed Agents动态工作流文档
Anthropic just published the 𝗳𝘂𝗹𝗹 𝗼𝗳𝗳𝗶𝗰𝗶𝗮𝗹 𝗱𝗼𝗰𝘀 for dynamic workflows in Managed…
不仅分享了官方新特性，还直接给出了可复用的复杂Prompt模板，结构清晰且有实测数据支撑，对开发者极具参考价值。
Anthropic just published the 𝗳𝘂𝗹𝗹 𝗼𝗳𝗳𝗶𝗰𝗶𝗮𝗹 𝗱𝗼𝗰𝘀 for dynamic workflows in Managed Agents. They planted 70 bugs in a 116K-line codebase. Across three runs, a single agent found 14, 15 and 27. A Managed Agents workflow ran three times and 𝗳𝗼𝘂𝗻𝗱 𝟲𝟲 𝗲𝘃𝗲𝗿𝘆 𝘁𝗶𝗺𝗲. You can have Opus 5.5 follow the docs to go through your project, pick the job most worth splitting across many agents, and run it as a dynamic workflow in Claude Code.
Anthropic 刚刚发布了托管代理（Managed Agents）中动态工作流的完整官方文档。他们在 11.6 万行代码库中植入了 70 个漏洞。在三次运行中，单个代理分别发现了 14、15 和 27 个漏洞。而一个托管代理工作流运行了三次，每次都发现了 66 个漏洞。你可以让 Opus 5.5 按照文档指引遍历你的项目，挑选出最值得拆分为多个代理的任务，并在 Claude Code 中以动态工作流的形式运行它。
In a dynamic workflow, Claude writes the plan first, hands the work to many agents in phases, then combines the results. Managed Agents bills at API rates. Claude Code has dynamic workflows too, on 𝗲𝘃𝗲𝗿𝘆 𝗽𝗮𝗶𝗱 𝗽𝗹𝗮𝗻, and they count toward your plan's usage (on Pro, turn them on in /config first). A single run can send out up to 1,000 agents, and every one of them uses your limits, so start with one directory. /workflows shows how many tokens each agent has used, and you can stop the run at any time.
在动态工作流中，Claude 首先编写计划，分阶段将工作分配给多个代理，然后汇总结果。托管代理按 API 费率计费。Claude Code 也支持动态工作流，适用于所有付费套餐，并且它们会计入套餐的使用量（在 Pro 套餐中，需先在 /config 中开启）。单次运行最多可发出 1,000 个代理，且每个代理都会消耗你的配额，因此建议先从单个目录开始。/workflows 会显示每个代理已使用的 token 数量，你也可以随时停止运行。
Send this prompt to Opus 5.5 👇
将此提示词发送给 Opus 5.5 👇
"Read these two docs: https://code.claude.com/docs/en/workflows https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration We'll run this as a dynamic workflow in Claude Code. From the second doc, only borrow how it splits and reviews the work.
"阅读以下两份文档： https://code.claude.com/docs/en/workflows https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration 我们将以动态工作流的形式在 Claude Code 中运行此任务。仅从第二份文档中借鉴其拆分工作和审查工作的方法。"
Then go through this project's code, issues and recent commits, and find three kinds of work that fit a dynamic workflow: 1. Checks that need to cover the whole repo, like hunting for bugs or finding missing permission checks. 2. Migrations that touch many files, like moving from JavaScript to TypeScript. 3. Fixes you have to repeat until they pass, like a pile of type-check errors. Stop if two rounds in a row make no progress. If the project doesn't have one of these, just say so. Don't force it.
接着浏览该项目的代码、问题（issues）和最近的提交（commits），找出三种适合动态工作流的工作类型： 1. 需要覆盖整个仓库的检查，例如查找漏洞或发现缺失的权限检查。 2. 涉及大量文件的迁移，例如从 JavaScript 迁移到 TypeScript。 3. 必须重复执行直到通过才能完成的修复，例如一堆类型检查错误。如果连续两轮没有任何进展，请停止。 如果该项目不具备上述任一类型，请直接说明。不要强行匹配。
Pick the one most worth doing and plan it on a single directory first, with no more than 10 agents for now. Give each piece of work one agent, and pair each piece with a second agent that reviews it and looks for what the first one missed. One agent failing must not fail the whole run; list anything that didn't get covered. If you pick a check, hand every issue it reports to another agent to pick apart, and only count the ones that hold up.
选择最值得执行的一项，并首先在单个目录上制定计划，目前代理数量不超过 10 个。为每项工作分配一个代理，并为每项工作配对第二个代理进行审查，查找第一个代理遗漏的内容。一个代理失败不得导致整个运行失败；列出任何未被覆盖的内容。如果你选择的是检查任务，将其报告的所有问题交给另一个代理逐一剖析，只保留那些经得起推敲的问题。
Show me in writing which one you picked, how many phases it has and how many agents it sends out. Don't start the workflow or change any files until I confirm."
书面告诉我你选择了哪一项，它包含多少个阶段，以及发送了多少个代理。在我确认之前，不要启动工作流或更改任何文件。"
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力