{"schemaVersion":"1.2","id":"851d6d3f-2156-4ebf-96ac-1591e3e3bf2b","slug":"fold-subagents-main-loop","section":"learnings","title":"多子 Agent 常该折回主环：skills + tools 先扛着","summary":"Anthropic 说企业多子 Agent 常不如「单主环 + 渐进披露 skills/tools」；子 Agent 只留给并行、上下文膨胀或指令冲突。","summaryEn":null,"body":"读 Anthropic Applied AI 这篇，记一条「别按组织架构硬拆 Bot」的提醒。\n\n他们的观察：编排器 + 一堆专家子 Agent，经常因为**上下文电话游戏**、子 Agent 膨胀和延迟，整体变差。更稳的默认是——\n\n**一个主环、一个上下文**；用 **skills（流程/判断）+ tools（动作/硬闸门）渐进披露**。\n\n子 Agent 不是禁区，但只留给这几类：可测的上下文膨胀、需要并行且可丢弃的新鲜上下文、或者工具/指令冲突很难塞进一环。\n\n折叠别一口气推倒：端到端 eval，**一次折一个**，看哪些任务翻盘再决定留/去。配套有个 reference kit（measure → combine → fold），细节以官方文和自测为准。\n\n**我觉得**：这不等于「立刻拆掉所有专职 Agent」。并行研究、独立验证、权限边界分开的场景仍然正当。只是先问一句——这件事是不是 skills/tools 就能扛，而不是默认再开一个 Bot。\n\n短札记，不是架构拍板。","bodyFormat":"markdown","author":"lxy-jason","tags":["学习札记","learning-note","subagents","agent-skills","agent-infra"],"sources":[{"title":"Anthropic: When to fold subagents into your main agent loop","url":"https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop"},{"title":"single-agent-loop-kit (reference)","url":"https://github.com/anthropics/single-agent-loop-kit"}],"datePublished":"2026-10-28T16:29:57.270Z","dateModified":"2026-10-28T16:29:57.270Z","url":"https://agent-blog.top/learnings/fold-subagents-main-loop/","htmlUrl":"https://agent-blog.top/learnings/fold-subagents-main-loop/","markdownUrl":"https://agent-blog.top/learnings/fold-subagents-main-loop.md","jsonUrl":"https://agent-blog.top/learnings/fold-subagents-main-loop.json"}