<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0"
     xmlns:atom="http://www.w3.org/2005/Atom"
     xmlns:content="http://purl.org/rss/1.0/modules/content/"
     xmlns:dc="http://purl.org/dc/elements/1.1/"
     xmlns:media="http://search.yahoo.com/mrss/">
  <channel>
    <title>Agent 系统设计解剖 on 熊鑫伟（cubxxw）· 公开的实践与思想档案</title>
    <link>https://cubxxw.com/zh/columns/agent-system-design/</link>
    <description>熊鑫伟的公开实践与思想档案，记录 AI 工作流、真实项目、工程方法、系统思维，以及关于个人成长、哲学与人生的长期思考。</description>
    <image>
      <title>熊鑫伟（cubxxw）· 公开的实践与思想档案</title>
      <url>https://cubxxw.com/assets/og-image.png</url>
      <link>https://cubxxw.com/</link>
    </image>
    <generator>Hugo -- gohugo.io</generator>
    <language>zh-cn</language>
    <lastBuildDate>Fri, 07 Aug 2026 19:44:00 +0800</lastBuildDate>
    <atom:link href="https://cubxxw.com/zh/columns/agent-system-design/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Agent 不是一种产品：十套系统如何重新分配控制权、状态、身份与副作用</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-synthesis/</link>
      <pubDate>Fri, 07 Aug 2026 19:43:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 19:43:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-synthesis/</guid>
      <description>基于十篇个案，以控制权、状态、执行身体、强制层和产品表面五个所有权问题，推导 bounded function、workflow、harness、event platform、gateway/computer 与 vertical organization 六种形态，并给出副作用宪法、五层停止协议和选型树。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">System Design</category>
      <category domain="tag">Data Processing</category>
      <category domain="tag">Security</category>
      <category domain="tag">Development</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/11-synthesis/agent-system-constitutional-route.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/11-synthesis/agent-system-constitutional-route.svg" medium="image"><media:description>Agent 系统设计的五个所有权问题、从意图到副作用与证据的控制权宪法之河，以及六种最小成立形态的决策图</media:description></media:content>
    </item>
    <item>
      <title>Conversation as Database：OpenHands 的无状态 Agent 与事件运行时</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-openhands/</link>
      <pubDate>Fri, 07 Aug 2026 19:40:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 19:40:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-openhands/</guid>
      <description>以 OpenHands SDK v1.41.0 为样本，沿事件链拆解无状态 step、Conversation 与 local/remote workspace。文章区分 event log、覆盖式 base state、派生 view 和外部世界，并以崩溃路径说明内部可回放为何仍不等于副作用恰好一次。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Data Processing</category>
      <category domain="tag">Monitoring</category>
      <category domain="tag">Security</category>
      <category domain="tag">Development</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/10-openhands/openhands-event-sourced-runtime.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/10-openhands/openhands-event-sourced-runtime.svg" medium="image"><media:description>OpenHands V1 从无状态 Agent step、追加式 typed event log、base state、派生 view 到 local/remote runtime 与副作用恢复协议的架构图</media:description></media:content>
    </item>
    <item>
      <title>多 Agent 辩论真的增加了信息吗：TradingAgents 的组织图与相关性风险</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-tradingagents/</link>
      <pubDate>Fri, 07 Aug 2026 18:50:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 18:50:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-tradingagents/</guid>
      <description>以 TradingAgents v0.3.1 与论文 v7 为样本，把分析师、牛熊辩手、交易员、风险声部和组合经理还原成真实 LangGraph。文章区分角色多样性与证据独立性，核验串行分析、共享模型、时间截点、checkpoint 和反思记忆，并以相关性、消融及外部执行成本检验多 Agent 辩论是否真正增加信息。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Data Processing</category>
      <category domain="tag">Monitoring</category>
      <category domain="tag">Security</category>
      <category domain="tag">Development</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/09-tradingagents/tradingagents-evidence-debate-graph.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/09-tradingagents/tradingagents-evidence-debate-graph.svg" medium="image"><media:description>TradingAgents 从时间截点证据、串行分析师、牛熊法庭、风险声部到组合经理研究输出的证据法庭架构图</media:description></media:content>
    </item>
    <item>
      <title>Instructions as Code：OpenMontage 如何用工件契约编排视频生产</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-openmontage/</link>
      <pubDate>Fri, 07 Aug 2026 18:20:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 18:20:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-openmontage/</guid>
      <description>以 OpenMontage 固定提交为样本，沿脚本、分镜、素材、剪辑、渲染与导出追踪规范工件，解释 coding assistant 如何用 YAML 和 Skill 驱动生产。文章拆开 tool、checkpoint、validator 与人工门，并以无音轨成片和付费素材崩溃检验恢复、成本、版权及发布边界。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Automation</category>
      <category domain="tag">Content Strategy</category>
      <category domain="tag">Development</category>
      <category domain="tag">Open Source</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/08-openmontage/openmontage-artifact-production-line.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/08-openmontage/openmontage-artifact-production-line.svg" medium="image"><media:description>OpenMontage 由 instruction 控制带、canonical artifact 胶片河和工具检查点验证带组成的视频生产架构图</media:description></media:content>
    </item>
    <item>
      <title>拒绝成为 Agent：TaxHacker 为什么把 LLM 压缩成一个函数</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-taxhacker/</link>
      <pubDate>Fri, 07 Aug 2026 17:55:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 17:55:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-taxhacker/</guid>
      <description>以 TaxHacker v0.8.5 为样本，追踪票据如何经过多模态抽取、provider fallback、动态 schema、表单校验、重复检查与人工保存，解释何时应把 LLM 限制成结构化函数。文章区分结构有效、事实正确与会计记录，并揭示 compatible 路径、前四页预览、缓存候选和金融数据的边界。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Document AI</category>
      <category domain="tag">Data Processing</category>
      <category domain="tag">Security</category>
      <category domain="tag">Development</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/07-taxhacker/taxhacker-bounded-ai.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/07-taxhacker/taxhacker-bounded-ai.svg" medium="image"><media:description>TaxHacker 将不可信票据经顺序 provider 抽取、schema 窄腰、原件核对、表单校验与显式保存变成 canonical transaction 的架构图</media:description></media:content>
    </item>
    <item>
      <title>OpenClaw 的常驻网关：连续性不是把所有消息塞进同一会话</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-openclaw/</link>
      <pubDate>Fri, 07 Aug 2026 17:40:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 17:40:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-openclaw/</guid>
      <description>以 OpenClaw v2026.7.1-2 为样本，追踪消息如何经过渠道准入、binding、agent 与 session key，解释跨渠道连续性的身份维度。文章拆开 Gateway、模型、workspace、sandbox 与设备节点，并用跨账户串线反例说明常驻 Agent 的权限、凭据、注入与撤销边界。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Security</category>
      <category domain="tag">Monitoring</category>
      <category domain="tag">Automation</category>
      <category domain="tag">Development</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/06-openclaw/openclaw-persistent-gateway.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/06-openclaw/openclaw-persistent-gateway.svg" medium="image"><media:description>OpenClaw 从互联网渠道经过常驻 Gateway、确定性 bindings、session key、per-agent enclave 到设备节点与外部副作用的路由图</media:description></media:content>
    </item>
    <item>
      <title>n8n 的确定性外骨骼：Queue、Worker 与副作用从不承诺 Exactly-Once</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-n8n/</link>
      <pubDate>Fri, 07 Aug 2026 17:25:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 17:25:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-n8n/</guid>
      <description>以 n8n 2.33.6 源码追踪一次执行如何跨主进程、Redis 队列、工作进程与 Postgres 流动，并拆开工作流图与 Agent 节点的控制边界。文章解释等待、重试、崩溃恢复、幂等键和对账各自解决什么，以及 queue mode 为什么只能扩大吞吐，不能承诺 exactly-once 副作用。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Automation</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Monitoring</category>
      <category domain="tag">Security</category>
      <category domain="tag">Development</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/05-n8n/n8n-deterministic-spine.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/05-n8n/n8n-deterministic-spine.svg" medium="image"><media:description>n8n 确定性 workflow 主干、受限 Agent 支线、人工与幂等闸门、外部回执对账以及 Redis worker Postgres queue-mode 车场</media:description></media:content>
    </item>
    <item>
      <title>给 Agent 一台电脑之后：Manus 如何把回复变成交付物</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-manus/</link>
      <pubDate>Fri, 07 Aug 2026 17:15:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 17:15:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-manus/</guid>
      <description>基于 Manus 2026-08-07 第一方资料，拆开临时 Sandbox、持久 Cloud Computer、Cloud Browser、Browser Operator 与 Wide Research。文章解释计算机如何把回复升级为可验收工件和持续服务，也说明登录态、外部副作用与恢复责任为何仍在 VM 边界外。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Product Strategy</category>
      <category domain="tag">Security</category>
      <category domain="tag">Automation</category>
      <category domain="tag">Cloud</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/04-manus/manus-computer-to-artifact.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/04-manus/manus-computer-to-artifact.svg" medium="image"><media:description>Manus task envelope、临时 Sandbox VM、browser files code software 流水线、artifact 验收、Wide Research 与本地 Browser Operator 边界</media:description></media:content>
    </item>
    <item>
      <title>Codex 的协议化内核：第二个 Surface 不该复制第一个 Agent</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-codex/</link>
      <pubDate>Fri, 07 Aug 2026 17:05:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 17:05:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-codex/</guid>
      <description>以 Codex 0.147.0、App Server 契约和 Rust core 为证据，拆开 Thread、Turn、Item 与 Submission、Event 两层协议。文章解释多种客户端如何共享 Agent loop、状态与治理，并厘清 approval、sandbox、持久化与产品 Surface 的边界。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Open Source</category>
      <category domain="tag">Development</category>
      <category domain="tag">Security</category>
      <category domain="tag">MCP</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/03-codex/codex-event-control-plane.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/03-codex/codex-event-control-plane.svg" medium="image"><media:description>Codex 多 Surface、App Server 公共契约、Submission 与 Event 双总线以及 approval 和 sandbox 副作用闸门</media:description></media:content>
    </item>
    <item>
      <title>Pi 的减法：最小 Agent Kernel 保留什么，责任又去了哪里</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-pi/</link>
      <pubDate>Fri, 07 Aug 2026 16:40:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 16:40:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-pi/</guid>
      <description>以 Pi v0.84.1 源码为证据，沿 agent loop、默认四工具、JSONL session tree 与 extension seam 反推最小 Agent kernel。文章审计 Pi 省略 Plan、MCP、subagent、permission 和 sandbox 后，安全、一致性和恢复责任的去向。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Open Source</category>
      <category domain="tag">Development</category>
      <category domain="tag">Security</category>
      <category domain="tag">Context Engineering</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/02-pi/pi-minimal-kernel.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/02-pi/pi-minimal-kernel.svg" medium="image"><media:description>Pi 最小 Agent kernel、可拆卸 extension 轨道、provider 与 host OS 信任边界以及下方 JSONL session tree</media:description></media:content>
    </item>
    <item>
      <title>Claude Code 的扩展语法：一条仓库修改怎样逐级获得控制</title>
      <link>https://cubxxw.com/zh/ai-agent/posts/agent-system-design-claude-code/</link>
      <pubDate>Fri, 07 Aug 2026 16:35:00 +0800</pubDate>
      <atom:updated>Fri, 07 Aug 2026 16:35:00 +0800</atom:updated>
      <dc:creator>熊鑫伟 (Xinwei Xiong)</dc:creator>
      <guid isPermaLink="true">https://cubxxw.com/zh/ai-agent/posts/agent-system-design-claude-code/</guid>
      <description>以 Claude Code v2.1.224 为证据边界，沿仓库修改控制链拆解 CLAUDE.md、Skill、MCP、Hook、permission、sandbox、subagent 与 worktree。文章区分软语境、能力接口、确定性拦截和文件隔离，说明模型何时拥有下一步，代码、操作系统、测试与人在何处接管后果。
</description>
      <category domain="tag">AI</category>
      <category domain="tag">Agent</category>
      <category domain="tag">Harness Engineering</category>
      <category domain="tag">Context Engineering</category>
      <category domain="tag">MCP</category>
      <category domain="tag">Security</category>
      <category domain="tag">Testing</category>
      <enclosure url="https://cubxxw.com/images/agent-system-series/01-claude-code/claude-code-progressive-control.svg" type="image/jpeg" length="0" />
      <media:content url="https://cubxxw.com/images/agent-system-series/01-claude-code/claude-code-progressive-control.svg" medium="image"><media:description>Claude Code 中模糊意图经过上下文、Agent loop、工具提案、权限、沙箱、工作树和测试，成为受约束仓库修改的控制河流</media:description></media:content>
    </item>
  </channel>
</rss>
