Anthropic 应用 AI 团队：在企业级部署对比中「a single agent with skills frequently outperformed（带技能的单智能体频繁跑赢）」子智能体设计——配套的 GitHub 迁移套件会「one at a time（逐个）」对着同一份评测折叠子智能体

“Anthropic 的应用 AI 团队发现，在多个企业级部署里，带技能的单智能体频繁跑赢子智能体设计——还发了一套在 GitHub 上逐次评测折叠的迁移套件。”

可用说法

• 2026 年 10 月 22 日（星期四，已按日历核算；页面自己的元信息行「Date October 22, 2026」与 JSON-LD datePublished 2026-10-22），Anthropic 在 claude.com 发布了 Diane Sarkis 的「Building agents: When to fold subagents into your main agent loop」。文章导语：「When building agents in Claude Platform, using fewer subagents in favor of progressively disclosed context and tools can improve quality while reducing cost and latency. Here’s when and how to migrate.（在 Claude Platform 上构建智能体时，少用子智能体、改用按需渐进披露的上下文与工具，可以在降低成本与延迟的同时提升质量。以下是何时以及如何迁移。）」框定：「Subagents – agents with their own context window, tools, and instructions, orchestrated by a coordinator – are a core component of agent design. They are effective at verification, parallel work, and modular permissioning. We use them in many of our own agents that we build using Claude Platform, for example our multi-agent research system.（子智能体——拥有自己上下文窗口、工具和指令、由协调者编排的智能体——是智能体设计的核心组件。它们在验证、并行工作和模块化权限管理上很有效。我们在许多用 Claude Platform 构建的自有智能体中都在使用它们，比如我们的多智能体研究系统。）」新发现：「However, recently the Applied AI team has found a pattern of enterprise agent deployments that have benefited from folding routers, planners, and domain subagents into the main agent loop.（不过，最近应用 AI 团队发现，一批企业级智能体部署从把路由器、规划器和领域子智能体并入主智能体循环中获益。）」头部发现：「In our comparisons across several enterprise deployments, a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task.（在我们对多个企业级部署的对比中，带技能的单智能体在质量上频繁跑赢「一个提示词包打天下」设计与子智能体设计，且单任务成本与延迟往往更低。）」客户示例，用信源自己的话：「one company found the dozen of specialist subagents within its customer-facing agent made a consistent on-brand experience difficult to maintain and added latency that hurt the user experience（一家公司发现其面向客户的智能体里那一打专家子智能体，让一致的品牌体验难以维持，还增加了伤害用户体验的延迟）」（信源自己的措辞，按原样引用）。熟悉演进模式的反馈：「most of this structure was doing work the model no longer needed done for it, and that one clear system prompt, 4-5 thin data tools, and firm output guardrails would cover most of the use case.（这些结构大部分在做模型不再需要别人替它做的活；一条清晰的系统提示词、4-5 个轻量数据工具加严格的输出护栏，就能覆盖大部分用例。）」边界：「To be clear, subagents still have their place, including at the research frontier.（明确地说，子智能体仍有其位置，包括在研究前沿。）」——在三种情况下保留：单智能体循环「hits measurable performance degradation from context bloat（因上下文膨胀出现可衡量的性能退化）」、「Multiple tasks need to be tackled in parallel with fresh context windows that can be discarded once the task is done（多个任务需要用全新的、用完即可丢弃的上下文窗口并行处理）」、或「A single agent struggles to select appropriate tools or there are conflicting instructions in the system prompt.（单智能体难以选出合适工具，或系统提示词里指令互相冲突。）」组织架构陷阱：「But too often we see enterprises give each team or problem type its own subagent, with the agent architecture becoming a copy of their org chart (Conway’s law at work).（但我们太常见到企业给每个团队或问题类型配一个子智能体，智能体架构变成了组织架构图的复制品（康威定律在起作用）。）」失败模式：「Subagent sprawl: Every team ships its own subagent, so companies end up with hundreds of overlapping subagents with no enforced standards and no specific owner.（子智能体蔓延：每个团队都发自己的子智能体，公司最后落得几百个互相重叠、没有统一标准、也没有明确归属者的子智能体。）」；「The Frankenprompt: Every team adds a sentence to the system prompt, resulting in conflicting rules and no consistent voice.（弗兰肯提示词：每个团队往系统提示词里加一句，结果规则互相冲突、语气无法统一。）」；延迟——「One customer that evaluated their multi-agent harness against a single agent loop saw a roughly 2x faster runtime with the latter.（一位客户把多智能体框架与单智能体循环做了对比评估，后者运行快了约 2 倍。）」模型能力驱动：「starting with Opus 4.5, models became capable of managing far more in a single loop than most teams expected. You cannot turn off thinking on Opus 5.5, and Fable 5.1, and recent models increasingly narrate their progress and check their work as they go, which makes some harness elements redundant.（从 Opus 4.5 开始，模型在单个循环里能管理的复杂度远超多数团队的预期。Opus 5.5 和 Fable 5.1 上无法关闭思考，而且新近的模型越来越会在过程中叙述进度、自查工作，这让框架的某些环节变得多余。）」以及「in migrations we have helped customers with, rigid multi-agent structures were often what stopped the new model from doing more of the work（在我们协助客户做的迁移中，僵硬的多智能体结构常常正是阻碍新模型承担更多工作的东西）」。提示词裁剪数据：「we reduced Claude Code’s system prompt for more advanced models by over 80% which included cuts related to progressively disclosing skills and tools.（我们把 Claude Code 面向更高级模型的系统提示词削减了 80% 以上，其中就包括与按需渐进披露技能和工具相关的删减。）」设计门槛：「The single agent loop design requires discipline: a carefully managed system prompt, skill and tool descriptions that trigger at the right times, and deliberate handling of context windows, memory, and note-taking.（单智能体循环设计需要纪律：精心管理的系统提示词、能在恰当时机触发的技能与工具描述，以及对上下文窗口、记忆和记笔记的刻意处理。）」反对盲目合并的反例：「When we tested this on a sample support agent, folding one subagent into another lowered latency, but also broke a process that had been passing. The main agent was holding the order lookup tool and answering on its own, without asking the shipping specialist for the delivery date.（我们在一个示例客服智能体上测试时，把一个子智能体并进另一个虽然降低了延迟，却也破坏了一个原本通过的流程。主智能体自己攥着订单查询工具直接作答，没有去问物流专家要送达日期。）」——附评测规则「run each task more than once, since a break like this one only showed up in one run out of two.（每个任务要跑不止一次，因为这样的破坏只在两次运行中的一次里露头。）」工具数阈值：「once you exceed 30-50 tool definitions in the system prompt, tool choice tends to degrade（一旦系统提示词里的工具定义超过 30-50 个，工具选择往往开始退化）」。迁移套件：「Our migration kit chains these skills together in one guided workflow for assessing and moving multi-agent systems to a single agent loop. It includes a demo you can try before running it on the real thing（我们的迁移套件把这些技能串成一个引导式工作流，用于评估多智能体系统并将其迁到单智能体循环。它带一个演示，你可以在动真格之前先试）」——链接指向 github.com/anthropics/single-agent-loop-kit。结尾：「As models get more capable and agents get longer running, more of your subagents can be skills and tools. You don’t have to make that move in one step, though. Build a small end-to-end eval, fold in one subagent, and keep it as a subagent only if the folded version scores worse, then rinse and repeat.（随着模型越来越强、智能体运行越来越长，你的子智能体可以有更多变成技能和工具。不过不必一步到位。先建一个小的端到端评测，并入一个子智能体，只有当并入版得分更差时才把它留作子智能体，然后如此往复。）」按套件自己的 README（本次运行从仓库 master 分支抓取）：「A starter kit for moving a multi-agent system (a router or orchestrator in front of specialist subagents) toward one agent loop that loads skills and tools on demand.（一套入门套件，用于把多智能体系统（专家子智能体前面的路由器或编排器）迁往按需加载技能和工具的单智能体循环。）」横幅：「Reference implementation. Not maintained and not accepting contributions.（参考实现。不维护、不接受贡献。）」「The kit does not assume a single loop always wins. It measures. Every change is scored on the same eval before and after, and a subagent only goes away when the numbers say its replacement holds.（套件不假定单循环永远赢。它靠测量说话。每一次改动都在同一个评测上做前后打分，只有数字说明替代方案站得住，子智能体才会被移走。）」三个步骤「Measure」「Combine」「Fold」——「Turn each remaining subagent into a skill and tools in the main agent, one at a time, and compare.（把其余每个子智能体逐个变成主智能体里的技能和工具，并做对比。）」「examples/support-desk is a small, runnable support agent with a router and six subagents（examples/support-desk 是一个可运行的小型客服智能体，带一个路由器和六个子智能体）」以及无评分警句「Without a score, a migration is a matter of opinion.（没有评分，迁移就只是各说各话。）」按 The Decoder（Matthias Bastian，datePublished 2026-10-25T09:12:10+00:00）——把这篇东西推向大众读者的报道：「Anthropic recommends that companies evaluate whether they can consolidate sub-agents into a single agent that loads context on demand.（Anthropic 建议企业评估能否把子智能体合并成一个按需加载上下文的单一智能体。）」「In one comparison, Anthropic says the single-agent architecture ran about twice as fast as the multi-agent setup.（在一次对比中，Anthropic 说单智能体架构的运行速度约为多智能体方案的两倍。）」以及「Anthropic warns against merging blindly, since in one test with a support agent, folding in a sub-agent did reduce latency but broke a previously working process.（Anthropic 警告不要盲目合并——在一次客服智能体测试中，并入子智能体确实降低了延迟，却也破坏了一个原本正常工作的流程。）」