{"id":"2c2e994b-5e03-47d7-91d1-972c20d17d1e","entityType":"agent","slug":"clawhub-gechengling-ai-agent-orchestration-advisor","name":"AI Agent Orchestration Advisor","canonicalUrl":"https://www.xpersona.co/agent/clawhub-gechengling-ai-agent-orchestration-advisor","canonicalPath":"/agent/clawhub-gechengling-ai-agent-orchestration-advisor","generatedAt":"2026-10-10T21:49:50.582Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-10T18:27:20.868Z","emptyReason":null},"description":"Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework. AI-powered multi-agent framework comparison and selection assistant — analyze use cases, compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体, 智能体编排, 框架对比, 智能体架构. Skill: AI Agent Orchestration Advisor Owner: gechengling Summary: Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework. AI-powered multi-agent framework comparison and selection assistant — analyze use cases, compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture recommendat","descriptionLabel":"Technical summary","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 1.3K downloads reported by the source. Last updated 10/10/2026.","installCommand":"clawhub skill install s17ewqc4f2s6gpcbm88hy7fgvn85kg1g:ai-agent-orchestration-advisor","sourceUrl":"https://clawhub.ai/gechengling/ai-agent-orchestration-advisor","homepage":"https://clawhub.ai/gechengling/skills/ai-agent-orchestration-advisor","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/gechengling/ai-agent-orchestration-advisor","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/gechengling/skills/ai-agent-orchestration-advisor","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":62,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any "},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-10T18:27:20.868Z","emptyReason":null},"protocols":[{"protocol":"OPENCLEW","label":"OpenClaw","status":"self-declared","notes":"Declared in the public agent profile."}],"capabilities":[],"verifiedCount":0,"selfDeclaredCount":1,"capabilityMatrix":{"rows":[{"key":"OPENCLEW","type":"protocol","support":"unknown","confidenceSource":"profile","notes":"Listed on profile"}],"flattenedTokens":"protocol:OPENCLEW|unknown|profile"}},"adoption":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-10T18:27:20.868Z","emptyReason":null},"stars":null,"forks":null,"downloads":1296,"packageName":null,"latestVersion":"3.0.2","tractionLabel":"1.3K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-10T18:27:20.810Z","emptyReason":null},"lastUpdatedAt":"2026-10-10T18:27:20.868Z","lastCrawledAt":"2026-10-10T18:27:20.810Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-11T18:27:20.810Z","lastVerifiedAt":null,"highlights":[{"version":"3.0.2","createdAt":"2026-10-09T05:20:00.414Z","changelog":"3.0.2: 新增数据最小化声明与执行边界(不安装/不运行/不部署框架)；收窄触发词；生态动态更新至2026-10-09；各步骤增补示例与表格维度","fileCount":3,"zipByteSize":14108},{"version":"3.0.1","createdAt":"2026-09-13T15:04:40.696Z","changelog":"内容增强：新增生态与合规动态表（框架成熟度/多智能体协作/SDK竞争/协议生态/长上下文/AI治理/可观测性）及两则解读示例；新增需求采集表与框架对比表维度扩展（可观测性、成本模型）并补充3个框架、新增选型决策表；新增编排模式对照表、人工复核卡点表、有状态骨架代码与交付物清单；新增评测维度、护栏清单、成本容量测算；生产检查清单扩充至12项；补充两则中文交互示例；修复正文重复段落","fileCount":3,"zipByteSize":10305},{"version":"3.0.0","createdAt":"2026-05-25T11:51:47.265Z","changelog":"- Major version update with expanded 2026 technical landscape coverage. - Enhanced framework comparison details for LangGraph v1.0, CrewAI v1.10, Claude Agent SDK, and OpenAI Agents SDK, including new performance metrics and cost benchmarks. - Added evaluation of MCP (Model Context Protocol) ecosystem growth and coverage. - Included guidance for LLM long-context selection with scenario-based recommendations. - Updated production checklist and starter code examples for contemporary multi-agent frameworks. - Improved support for enterprise requirements and rapidly evolving agent protocol integrations.","fileCount":3,"zipByteSize":5014},{"version":"1.0.1","createdAt":"2026-05-15T23:26:48.098Z","changelog":"No changes detected in this version. - No file or documentation changes present between previous and current versions. - Functionality and feature set remain unchanged.","fileCount":2,"zipByteSize":3848},{"version":"1.0.0","createdAt":"2026-05-15T14:11:55.118Z","changelog":"Initial release — expert advisor for multi-agent AI framework selection and architecture. - Compares top orchestration frameworks (LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Claude Agent SDK, Strands Agents) based on the latest 2026 ecosystem. - Analyzes user requirements and generates tailored framework comparisons, architecture diagrams, and starter Python scaffolds. - Includes current technical benchmarks: state management, tool accuracy, cost, long-context capabilities, and enterprise features. - Offers workflow from use case assessment to production checklist, covering error handling, observability, and scaling. - Supports key developer needs: team migrations, complex stateful workflows, and cross-framework trade-offs.","fileCount":2,"zipByteSize":3847}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s17ewqc4f2s6gpcbm88hy7fgvn85kg1g:ai-agent-orchestration-advisor","setupComplexity":"low","setupSteps":["Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.","Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data."],"contract":{"contractStatus":"missing","authModes":[],"requires":[],"forbidden":[],"supportsMcp":false,"supportsA2a":false,"supportsStreaming":false,"inputSchemaRef":null,"outputSchemaRef":null,"dataRegion":null,"contractUpdatedAt":null,"sourceUpdatedAt":null,"freshnessSeconds":null},"invocationGuide":{"preferredApi":{"snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-10T21:49:50.580Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-gechengling-ai-agent-orchestration-advisor/trust"}},"reliability":{"evidence":{"source":"runtime-metrics","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No trust, reliability, or runtime telemetry is available."},"trust":{"status":"unavailable","handshakeStatus":"UNKNOWN","verificationFreshnessHours":null,"reputationScore":null,"p95LatencyMs":null,"successRate30d":null,"fallbackRate":null,"attempts30d":null,"trustUpdatedAt":null,"trustConfidence":"unknown","sourceUpdatedAt":null,"freshnessSeconds":null},"decisionGuardrails":{"doNotUseIf":["Contract metadata is missing or unavailable for deterministic execution."],"safeUseWhen":[],"riskFlags":["missing_or_unavailable_contract","trust_data_unavailable","schema_references_missing"],"operationalConfidence":"low"},"executionMetrics":{"observedLatencyMsP50":null,"observedLatencyMsP95":null,"estimatedCostUsd":null,"uptime30d":null,"rateLimitRpm":null,"rateLimitBurst":null,"lastVerifiedAt":null,"verificationSource":null},"runtimeMetrics":{"successRate":null,"avgLatencyMs":null,"avgCostUsd":null,"hallucinationRate":null,"retryRate":null,"disputeRate":null,"p50Latency":null,"p95Latency":null,"lastUpdated":null}},"benchmarks":{"evidence":{"source":"no-benchmark-data","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No benchmark suites or observed failure patterns are available."},"suites":[],"failurePatterns":[]},"artifacts":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"high","updatedAt":"2026-10-10T18:27:20.868Z","emptyReason":null},"readme":"Skill: AI Agent Orchestration Advisor\n\nOwner: gechengling\n\nSummary: Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework. AI-powered multi-agent framework comparison and selection assistant — analyze use cases, compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体, 智能体编排, 框架对比, 智能体架构.\n\nTags: ai-agent-orchestration-advisor:3.0.2, latest:3.0.2\n\nVersion history:\n\nv3.0.2 | 2026-10-09T05:20:00.414Z | user\n\n3.0.2: 新增数据最小化声明与执行边界(不安装/不运行/不部署框架)；收窄触发词；生态动态更新至2026-10-09；各步骤增补示例与表格维度\n\nv3.0.1 | 2026-09-13T15:04:40.696Z | user\n\n内容增强：新增生态与合规动态表（框架成熟度/多智能体协作/SDK竞争/协议生态/长上下文/AI治理/可观测性）及两则解读示例；新增需求采集表与框架对比表维度扩展（可观测性、成本模型）并补充3个框架、新增选型决策表；新增编排模式对照表、人工复核卡点表、有状态骨架代码与交付物清单；新增评测维度、护栏清单、成本容量测算；生产检查清单扩充至12项；补充两则中文交互示例；修复正文重复段落\n\nv3.0.0 | 2026-05-25T11:51:47.265Z | auto\n\n- Major version update with expanded 2026 technical landscape coverage.\n- Enhanced framework comparison details for LangGraph v1.0, CrewAI v1.10, Claude Agent SDK, and OpenAI Agents SDK, including new performance metrics and cost benchmarks.\n- Added evaluation of MCP (Model Context Protocol) ecosystem growth and coverage.\n- Included guidance for LLM long-context selection with scenario-based recommendations.\n- Updated production checklist and starter code examples for contemporary multi-agent frameworks.\n- Improved support for enterprise requirements and rapidly evolving agent protocol integrations.\n\nv1.0.1 | 2026-05-15T23:26:48.098Z | auto\n\nNo changes detected in this version.\n\n- No file or documentation changes present between previous and current versions.\n- Functionality and feature set remain unchanged.\n\nv1.0.0 | 2026-05-15T14:11:55.118Z | auto\n\nInitial release — expert advisor for multi-agent AI framework selection and architecture.\n\n- Compares top orchestration frameworks (LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Claude Agent SDK, Strands Agents) based on the latest 2026 ecosystem.\n- Analyzes user requirements and generates tailored framework comparisons, architecture diagrams, and starter Python scaffolds.\n- Includes current technical benchmarks: state management, tool accuracy, cost, long-context capabilities, and enterprise features.\n- Offers workflow from use case assessment to production checklist, covering error handling, observability, and scaling.\n- Supports key developer needs: team migrations, complex stateful workflows, and cross-framework trade-offs.\n\nArchive index:\n\nArchive v3.0.2: 3 files, 14108 bytes\n\nFiles: skill-card.md (2112b), SKILL.md (27585b), _meta.json (149b)\n\nFile v3.0.2:SKILL.md\n\n---\r\nname: AI Agent Orchestration Advisor\r\ndescription: >\r\n  Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework.  AI-powered multi-agent framework comparison and selection assistant — analyze use cases,\r\n  compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture\r\n  recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph,\r\n  CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework\r\n  comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体,\r\n  智能体编排, 框架对比, 智能体架构.\r\nversion: \"3.0.2\"\r\n---\r\n\r\n# AI Agent Orchestration Advisor / 多智能体编排选型顾问\r\n\r\n> Your expert co-pilot for designing, selecting, and implementing multi-agent AI systems.\r\n> 面向多智能体系统的选型、架构设计与落地实施：从需求澄清到框架对比、架构拓扑、起步代码、评测护栏与成本测算。\r\n\r\n## What This Skill Does\r\n\r\nIn 2026, the agentic AI ecosystem exploded — LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Claude Agent SDK, and Strands Agents all compete for developer mindshare. Picking the wrong framework wastes weeks. This skill helps you:\r\n\r\n- **Choose the right framework** for your specific use case (workflow complexity, state management, team size, hosting requirements)\r\n- **Generate architecture diagrams** and data flow specs for multi-agent systems\r\n- **Produce starter code** scaffolds (Python) for the chosen framework\r\n- **Analyze trade-offs** across orchestration patterns (hierarchical, sequential, parallel, event-driven)\r\n- **Design evaluation and guardrails** before go-live, including human-review checkpoints\r\n- **Estimate cost and capacity** per run and per month\r\n- **Debug and optimize** existing multi-agent implementations\r\n\r\n## Trigger Words / 触发词\r\n\r\n**English Triggers:** multi-agent framework comparison, agent orchestration design, LangGraph vs CrewAI, agent architecture topology, agent evaluation and guardrails, agent cost estimation, agent state persistence debugging\r\n\r\n**English Non-Triggers:** single prompt writing, general chatbot building, model fine-tuning, RAG pipeline basics unrelated to agents, web scraping scripts, generic Python debugging\r\n\r\n**中文触发词（须落在多智能体选型或设计任务上才触发）：** 多智能体框架怎么选 / 智能体编排模式有哪些 / 有状态智能体怎么做断点恢复 / 多智能体评测指标怎么定 / 智能体护栏怎么设计 / 多智能体成本怎么测算 / 智能体循环失控怎么处理\r\n\r\n**不触发清单：** 单条提示词优化、通用大模型问答、模型微调与训练、普通脚本编写、纯项目管理咨询\r\n\r\n**路由判定三步**：① 任务是否涉及两个及以上 Agent 的协作、编排或框架选型？② 是否需要输出架构、代码脚手架、评测或成本结论？③ 前两步均为是才启用本技能；单 Agent 场景转提示词工程或应用开发类技能。\r\n\r\n## Target Users / 适用人群\r\n\r\n\r\n- AI engineers building production agent systems\r\n- Data scientists exploring agentic automation\r\n- Product managers scoping agent-based features\r\n- Developers migrating from single LLM to multi-agent pipelines\r\n- **金融行业科技与数据团队**：需同时满足业务交付与合规留痕要求\r\n\r\n---\r\n\r\n## Ecosystem & Compliance Updates [2026-10-09] / 生态与合规动态\r\n\r\n| 类型 | 内容摘要 | 对选型的影响 | 落地动作 | 优先级 |\r\n|-----|---------|-----------|---------|-------|\r\n| 框架成熟度 | LangGraph 状态机路线趋于稳定：状态持久化/长期记忆/错误恢复成为基线能力，企业级部署支持容器编排扩缩容 | 复杂有状态工作流首选评估对象 | 以\"状态恢复是否可验证\"作为 POC 验收项 | 高 |\r\n| 多智能体协作 | CrewAI 路线强化角色化协作与并行任务编排，连接器生态扩张 | 角色清晰、流程线性的场景上手成本低 | 先用 2-3 个角色做最小闭环 | 高 |\r\n| SDK 竞争 | Claude Agent SDK / OpenAI Agents SDK 在工具调用与上下文利用上各有侧重 | 与既有模型生态绑定程度决定迁移成本 | 以\"模型可替换性\"评估锁定风险 | 高 |\r\n| 协议生态 | MCP 生态持续扩张，企业内部 MCP 注册表逐步成为基础设施 | 工具接入标准化，减少定制胶水代码 | 建立内部工具注册与权限清单 | 中 |\r\n| 互操作协议 | Agent 间互操作协议（如 A2A 方向）推进，跨系统协作成为议题 | 跨部门/跨厂商协作场景需预留适配层 | 关键接口做抽象层，避免直连 | 中 |\r\n| 长上下文 | 主流模型上下文窗口差异显著，长文档场景成本差异被放大 | 金融长文档场景需按\"分片 + 检索\"而非全量投喂 | 以 10 万字符文档做成本对照测试 | 中 |\r\n| AI 治理 | 金融机构 AI 应用需满足安全开发与合规使用要求，涉及场景清单、数据来源与人工复核 | Agent 系统需自带留痕与复核环节 | 上线清单中加入治理条目 | 高 |\r\n| 可观测性 | 评测与追踪工具成为生产必备 | 无追踪能力的框架运维成本高 | 选型时把追踪能力纳入打分 | 中 |\r\n| 上下文工程 | 上下文压缩、剪枝与长期记忆分层成为降本主线，不再是简单扩大窗口 | 长任务成本结构改变 | 把「每步上下文净增量」纳入成本测算项 | 高 |\r\n| 沙箱与隔离 | 代码执行类工具普遍要求沙箱化与资源限额 | 涉及代码执行的 Agent 需额外隔离设计 | 工具清单中标注哪些必须在沙箱内运行 | 高 |\r\n\r\n> **数据截止**: 2026-10-09 | 来源：各框架官方文档与公开版本说明、协议公开资料、行业公开信息\r\n> **声明**: 以上为生态观察与技术选型参考，框架版本与能力请以官方最新发布为准\r\n\r\n**动态解读示例（两类高频场景）**\r\n\r\n- **场景A｜框架锁定风险被低估**：团队为快速交付直接把业务逻辑写入某 SDK 的专有抽象层，半年后需要更换模型供应商 → 迁移成本约等于重写。**改进动作**：把\"模型调用、工具定义、记忆存储\"三类接口做薄抽象层；POC 阶段就验证一次模型替换（换供应商跑同一套评测集）。\r\n- **场景B｜缺失追踪导致问题无法定位**：多智能体跑出错误结论，但未记录中间步骤与工具调用参数 → 无法复盘是检索错、工具错还是生成错。**改进动作**：选型时将\"每步输入输出可导出、可回放\"列为硬性要求，并在 POC 中导出一次完整调用链。\r\n\r\n- **场景C｜并行分支的结果合并无一致性处理**：三个子 Agent 并行取数后直接拼接输出，未定义冲突消解规则 → 同一问题两次运行结果不一致，且无法判断哪个对。**改进动作**：并行分支后固定一个「汇总与冲突消解」节点，明确冲突时的优先级（可信来源优先 / 时间戳优先 / 转人工），并在评测集中加入一致性用例（同输入跑 10 次比对）。\r\n\r\n---\r\n\r\n## Data Minimization & Execution Boundary / 数据最小化与执行边界\r\n\r\n**数据最小化前置声明（使用本技能前请先执行）**\r\n\r\n1. 描述需求时不要粘贴真实业务数据、客户信息、内部接口地址与账号口令；用字段化描述替代（如「一批合同文本，含甲乙双方与金额」）。\r\n2. 起步代码中的 API Key、数据库连接串一律使用占位符（如 `os.environ[\"...\"]`），不得写入明文凭据。\r\n3. 需要评估真实数据时，先做脱敏与抽样，且样本不得离开机构内网。\r\n4. 本技能不安装、不运行、不部署任何框架；不访问用户的代码仓库、终端或云端环境；不发起任何外部网络调用。\r\n5. 生成的代码脚手架与配置如需落盘或提交，须先预览全文、确认无凭据与内网地址残留，再由责任人按机构变更流程办理。\r\n\r\n**代码块性质与执行边界**\r\n\r\n| 内容 | 性质 | 谁来执行 |\r\n|------|------|---------|\r\n| Step 4 的 CrewAI 示例 | 可运行脚手架，需在用户本地安装依赖后执行 | 使用者在自己的环境中运行，本技能不执行 |\r\n| Step 4.1 的 LangGraph 骨架 | 片段示例，省略了图构建与检查点配置 | 使用者补全后运行，本技能不执行 |\r\n| 框架对比表中的能力标注 | 生态观察结论，随版本变化 | 使用者以官方文档复核后采用 |\r\n| 部署检查清单 | 上线前的自查项 | 使用者的工程与合规团队 |\r\n\r\n**硬边界**：不执行代码、不安装依赖、不读写用户文件系统、不发起网络请求、不访问凭据、不代为部署上线。\r\n\r\n\r\n## Step 1 — Understand the Use Case\r\n\r\nAsk the user to describe:\r\n\r\n- The task or workflow to automate (e.g., \"research + summarize + post\")\r\n- Number of distinct roles/agents needed\r\n- State persistence requirements (ephemeral vs. persistent)\r\n- Hosting preference (cloud / local / serverless)\r\n- Team's programming experience\r\n\r\n### 1.1 Requirement Intake Table / 需求采集表\r\n\r\n| 维度 | 关键问题 | 为什么重要 | 常见误判 | 判定依据（可用于追问） |\r\n|-----|---------|-----------|---------|---|\r\n| 任务形态 | 线性流程、分支决策还是开放探索？ | 决定编排模式（顺序/图/事件驱动） | 把线性流程做成多智能体 | 能否画出只有单向边的流程图？能则偏线性 |\r\n| 角色数量 | 真正需要几个不同职责？ | 角色越多，通信成本与失败点越多 | 为\"看起来强大\"堆角色 | 两个角色职责是否可被同一个人完成？能则合并 |\r\n| 状态需求 | 是否需要跨会话记忆与断点恢复？ | 决定是否必须有持久化检查点 | 用外部数据库硬凑 | 中断后是否需要从断点继续？需要则必须持久化 |\r\n| 延迟预算 | 单次任务可接受多少秒？ | 多跳调用会线性放大延迟 | 忽略并行化的可行性 | 用户会在页面前等待还是异步取结果？ |\r\n| 成本上限 | 单次任务可接受成本？ | 角色数与上下文窗口直接放大成本 | 按\"每个 Token 单价\"估算而非按任务 | 单任务成本是否高于人工成本的 30%？ |\r\n| 部署环境 | 公有云、私有化还是一体机？ | 部分框架/服务在私有化下受限 | 忽略内部网络与数据出域限制 | 是否有数据不出内网的硬性约束？ |\r\n| 团队技能 | 团队熟悉 Python 异步与状态管理吗？ | 决定学习曲线是否可承受 | 低估调试成本 | 团队能否独立排查异步死锁与状态污染？ |\r\n| 合规要求 | 是否需要留痕、复核、可解释？ | 金融场景为硬性要求 | 上线前才补 | 是否会被审计或监管调阅？ |\r\n\r\n---\r\n\r\n## Step 2 — Framework Shortlist & Comparison\r\n\r\n### 2.1 Comparison Table / 框架对比\r\n\r\n| Framework | Best For | State Mgmt | Learning Curve | Hosting | Observability | 成本模型 | 锁定风险 |\r\n|-----------|----------|------------|----------------|---------|--------------|--------|---|\r\n| LangGraph | Complex stateful workflows | ✅ Built-in | Medium | Any | 强（图级追踪） | 按节点调用计费 | 图抽象与框架绑定，迁移需重写节点 |\r\n| CrewAI | Role-based team simulations | Partial | Low | Any | 中 | 按角色任务计费 | 角色与任务模型绑定，深度定制时受限 |\r\n| AutoGen/AG2 | Conversational agent loops | External | Medium | Any | 中 | 按轮次计费 | 会话循环模型绑定，转向图结构成本较高 |\r\n| OpenAI Agents SDK | OpenAI ecosystem, handoffs | Built-in | Low | Cloud-first | 中 | 按调用计费 | 与厂商模型能力强绑定，换供应商需改造 |\r\n| Claude Agent SDK | Anthropic native, tool use | Built-in | Low | Cloud-first | 中 | 按调用计费 | 同上，且工具协议为厂商私有 |\r\n| Strands Agents | AWS/Bedrock integration | External | Medium | AWS | 依赖云侧 | 按云服务计费 | 与云服务强绑定，私有化受限 |\r\n| LlamaIndex Workflows | 检索增强类流程 | Built-in | Low | Any | 中 | 按调用计费 | 检索组件绑定，非检索类流程优势不明显 |\r\n| Semantic Kernel | .NET/多语言企业栈 | External | Medium | Any | 中 | 按调用计费 | 企业栈绑定，生态插件相对少 |\r\n| 低代码平台（Dify/Coze 类） | 内部工具、快速验证 | Built-in | Very Low | SaaS/私有 | 弱至中 | 按订阅计费 | 平台绑定，复杂逻辑与私有化受限 |\r\n\r\n### 2.2 Selection Decision Table / 选型决策表\r\n\r\n| 情形 | 首选 | 次选 | 理由 | POC 验收项 |\r\n|-----|------|------|------|---|\r\n| 有状态、需断点恢复的长流程 | LangGraph | LlamaIndex Workflows | 图结构与检查点是核心诉求 | 断电/中断后能从检查点恢复并继续 |\r\n| 角色清晰、线性协作 | CrewAI | AutoGen/AG2 | 上手快，角色定义直观 | 两角色最小闭环一次跑通且可回放 |\r\n| 需要与既有模型生态深度绑定 | 对应厂商 SDK | LangGraph | 官方 SDK 的工具与回退机制更完整 | 换一个模型供应商跑同一评测集结果不劣化 |\r\n| 强检索依赖（知识库问答） | LlamaIndex Workflows | LangGraph | 检索组件成熟 | 同题 10 次检索命中一致 |\r\n| 企业 .NET 技术栈 | Semantic Kernel | LangGraph | 与既有服务栈一致 | 与既有 .NET 服务互通，无胶水层堆积 |\r\n| 快速验证业务想法 | 低代码平台 | CrewAI | 一周内可出可用原型 | 一周内产出可被业务方使用的原型 |\r\n| 私有化 + 强合规留痕 | LangGraph | 自研轻量状态机 | 追踪与持久化可完全自控 | 调用链可完整导出至内网存储 |\r\n\r\n> **规则**：先用\"流程是否有状态 + 是否需要检索 + 部署限制\"三个问题筛掉大部分选项，再在剩余的 2-3 个里做 POC 对比。\r\n\r\n---\r\n\r\n## Step 3 — Architecture Recommendation\r\n\r\nOutput a recommended architecture including:\r\n\r\n- Agent topology (who calls whom)\r\n- Tool assignments per agent\r\n- Memory / state strategy\r\n- Human-in-the-loop checkpoints\r\n- Error handling & fallback patterns\r\n\r\n### 3.1 Orchestration Patterns / 编排模式\r\n\r\n| 模式 | 适用场景 | 优点 | 风险 | 典型实现 | 成本特征 |\r\n|-----|---------|------|------|---------|---|\r\n| 顺序（Sequential） | 步骤确定的流水线 | 易调试、成本可预测 | 缺乏动态调整能力 | CrewAI Process.sequential | 成本随步骤线性增长，最可预测 |\r\n| 层级（Hierarchical） | 需要主管分配与汇总 | 复杂任务可分工 | 主管节点易成瓶颈 | 主管 Agent + 子 Agent | 主管节点上下文膨胀，成本集中在汇总层 |\r\n| 并行（Parallel） | 多源独立取数/多方案比较 | 降低总延迟 | 结果合并需一致性处理 | 并发任务 + 汇总节点 | 总成本不一定下降，但墙钟时间显著下降 |\r\n| 事件驱动（Event-driven） | 长时任务、需外部触发 | 解耦、可扩展 | 状态追踪复杂 | 队列 + 状态机 | 空闲等待不计费，但长驻状态有存储开销 |\r\n| 评审循环（Critic loop） | 质量要求高的产出 | 自我纠错 | 循环失控导致成本飙升 | 生成 + 评审双角色 | 成本随轮次倍增，必须设上限 |\r\n\r\n**拓扑示例（层级 + 评审循环）**\r\n\r\n```\r\nCoordinator\r\n ├── Retriever Agent   → 工具：知识库检索\r\n ├── Analyst Agent     → 工具：结构化计算\r\n └── Writer Agent      → 工具：文档生成\r\n        ↑\r\n   Critic Agent ── 不通过则回退到对应节点（限 2 轮）\r\n```\r\n\r\n### 3.2 Human-in-the-Loop Checkpoints / 人工复核卡点\r\n\r\n| 卡点 | 触发条件 | 人工动作 | 留痕要求 | 未设卡点的后果 |\r\n|-----|---------|---------|---------|---|\r\n| 外发前 | 产出将对外使用 | 复核事实与表述 | 复核人、时间、版本 | 错误表述直接对外，追溯时无法定位复核人 |\r\n| 高金额/高风险决策 | 涉及资金或风险判断 | 双人复核 | 双人签字或系统记录 | 单人决策错误无第二道防线，且责任不清 |\r\n| 低置信输出 | 置信度低于阈值 | 转人工处理 | 标注不确定点 | 低置信结论被当作确定结论使用 |\r\n| 循环失控 | 评审循环超过预设轮数 | 中止并转人工 | 保留循环记录 | 循环空转持续消耗预算，且产出不可用 |\r\n\r\n---\r\n\r\n## Step 4 — Starter Code Generation\r\n\r\nGenerate a complete, runnable Python scaffold:\r\n\r\n```python\r\n# Example: CrewAI research + report pipeline\r\nfrom crewai import Agent, Task, Crew, Process\r\nfrom crewai_tools import SerperDevTool\r\n\r\nsearch_tool = SerperDevTool()\r\n\r\nresearcher = Agent(\r\n    role=\"Senior Research Analyst\",\r\n    goal=\"Uncover cutting-edge developments in {topic}\",\r\n    backstory=\"You are an expert researcher...\",\r\n    tools=[search_tool],\r\n    verbose=True\r\n)\r\n\r\nwriter = Agent(\r\n    role=\"Technical Writer\",\r\n    goal=\"Craft insightful, accurate reports from research\",\r\n    backstory=\"You transform raw research into executive summaries...\",\r\n    verbose=True\r\n)\r\n\r\nresearch_task = Task(\r\n    description=\"Research {topic} thoroughly...\",\r\n    agent=researcher,\r\n    expected_output=\"Bullet-point research findings\"\r\n)\r\n\r\nwrite_task = Task(\r\n    description=\"Write a 500-word report on the research findings\",\r\n    agent=writer,\r\n    expected_output=\"Polished report with sections\"\r\n)\r\n\r\ncrew = Crew(\r\n    agents=[researcher, writer],\r\n    tasks=[research_task, write_task],\r\n    process=Process.sequential,\r\n    verbose=True\r\n)\r\n\r\nresult = crew.kickoff(inputs={\"topic\": \"agentic AI in 2026\"})\r\n```\r\n\r\n### 4.1 Stateful Skeleton (LangGraph style) / 有状态骨架\r\n\r\n```python\r\nfrom typing import TypedDict\r\n\r\nclass State(TypedDict):\r\n    query: str\r\n    findings: list\r\n    draft: str\r\n    review_passed: bool\r\n    rounds: int\r\n\r\ndef route_after_review(state: State) -> str:\r\n    \"\"\"评审后的路由：通过则结束，未通过且轮次未超限则回退\"\"\"\r\n    if state[\"review_passed\"]:\r\n        return \"end\"\r\n    if state[\"rounds\"] >= 2:\r\n        return \"human_review\"      # 循环超限 → 转人工，避免无限循环\r\n    return \"revise\"\r\n\r\n# 关键点：\r\n# 1) 检查点持久化，支持中断恢复\r\n# 2) 循环轮次上限，防止成本失控\r\n# 3) 低置信输出走人工复核分支\r\n```\r\n\r\n### 4.2 交付物清单 / Deliverables Checklist\r\n\r\n- [ ] 架构拓扑图（节点、边、条件路由）\r\n- [ ] 每个 Agent 的职责边界与工具清单\r\n- [ ] 状态结构定义与持久化方案\r\n- [ ] 评审与回退规则（含轮次上限）\r\n- [ ] 人工复核卡点清单\r\n- [ ] 评测集与验收标准\r\n\r\n\r\n**模块4 示例（两条）**\r\n\r\n- **示例A｜四角色流程的骨架拆解**：需求为「自动生成行业简报，需可回溯、可复核」。骨架拆解：Retriever（工具：内网知识库检索，只读）+ Analyst（工具：结构化计算，无外部网络）+ Writer（工具：文档生成）+ Critic（无工具，仅评审）；状态结构含 `query / findings / draft / review_passed / rounds`；`route_after_review` 中 `rounds >= 2` 强制转 `human_review`；检查点持久化到内网库。**关键约束**：Critic 不通过时回退到对应节点而非从头重跑，避免重复计费。\r\n- **示例B｜起步代码的三处必改项**：拿到 CrewAI 脚手架后必改：① `SerperDevTool`（外网检索）改为内网检索工具或明确列入数据出域审批；② API Key 改为从环境变量读取，禁止硬编码；③ `verbose=True` 的日志改为结构化落盘并脱敏，避免原始业务文本进入日志。未改完不上测试环境。\r\n\r\n---\r\n\r\n## Step 5 — Evaluation & Guardrails / 评测与护栏\r\n\r\n### 5.1 Evaluation Dimensions / 评测维度\r\n\r\n| 维度 | 指标示例 | 采集方式 | 参考门槛 | 不达标时的处理 |\r\n|-----|---------|---------|---------|---|\r\n| 任务完成率 | 端到端成功比例 | 评测集回放 | ≥85% | 定位失败最集中的节点，先砍角色再调提示 |\r\n| 事实准确性 | 与可信来源一致的比例 | 抽样核对 | 关键字段 100% | 关键字段改走确定性抽取，不让模型自由生成 |\r\n| 工具调用正确率 | 参数与调用对象正确比例 | 调用日志 | ≥95% | 给工具加参数校验与重试，必要时收窄工具描述 |\r\n| 稳定性 | 重复运行结果一致性 | 同输入跑 10 次 | 结论一致 | 固定随机性来源（温度、检索 Top-K），再做对照 |\r\n| 延迟 | 端到端耗时 | 埋点 | 符合延迟预算 | 把串行改并行，或降低单跳上下文 |\r\n| 成本 | 单任务成本 | 用量统计 | 符合成本上限 | 先降上下文净增量，再谈降角色数 |\r\n| 可回放性 | 调用链可导出比例 | 追踪系统 | 100% | 补齐追踪埋点后再上线，否则不具备投产条件 |\r\n\r\n### 5.2 Guardrails / 护栏清单\r\n\r\n| 护栏 | 实现方式 | 检查点 | 失效表现 |\r\n|-----|---------|-------|---|\r\n| 工具白名单 | 仅注册允许的工具 | 上线前审核 | 出现未注册工具被调用 |\r\n| 输入输出过滤 | 敏感信息与不当内容拦截 | 每次调用 | 敏感字段出现在输出或日志中 |\r\n| 循环上限 | 迭代轮次与总调用数上限 | 运行时 | 单任务调用数远超预算，账单异常 |\r\n| 超时与熔断 | 单步与整体超时 | 运行时 | 任务长时间挂起占用资源 |\r\n| 权限最小化 | 工具按最小权限授权 | 上线前 | 工具可访问超出职责范围的数据 |\r\n| 数据出域控制 | 敏感数据不出内网 | 架构评审 | 内网数据出现在外部服务日志中 |\r\n| 全链路留痕 | 输入、输出、参数、人工动作 | 100% | 事后无法复盘某次结论的生成过程 |\r\n\r\n---\r\n\r\n## Step 6 — Cost & Capacity Estimation / 成本与容量测算\r\n\r\n| 项目 | 测算口径 | 示例（单任务） | 压缩手段 |\r\n|-----|---------|-------------|---|\r\n| 模型调用 | 各角色 Token 用量 × 单价 | 3 角色 × 8K Token | 减少角色数、降低每步上下文净增量 |\r\n| 检索调用 | 检索次数 × 单次成本 | 5 次 | 检索结果先摘要再入上下文，控制 Top-K |\r\n| 外部工具 | 第三方 API 调用次数 | 2 次 | 合并调用、加缓存，避免重复查询 |\r\n| 延迟 | 串行跳数 × 单跳耗时 | 4 跳 × 3 秒 | 串行改并行，缩短关键路径 |\r\n| 重试开销 | 预计重试率 | +15% | 先消除失败根因，再容忍重试 |\r\n\r\n**示例测算**：一个\"检索 + 分析 + 撰写 + 评审\"四角色流程，单任务约 4 跳、合计约 3 万 Token；若以固定单价估算，可通过\"降低角色数 + 并行取数 + 结果复用\"把单任务成本压缩 30%-50%。**测算应基于实测用量，而非按单价线性外推。**\r\n\r\n---\r\n\r\n## Step 7 — Production Checklist / 生产检查清单\r\n\r\n- [ ] Rate limiting & retry logic\r\n- [ ] Observability (LangSmith / Weights & Biases / custom logging)\r\n- [ ] Secrets management (never hardcode API keys)\r\n- [ ] Cost estimation per run\r\n- [ ] Human review gates for high-stakes outputs\r\n- [ ] 状态持久化与中断恢复验证\r\n- [ ] 循环轮次与总调用数上限\r\n- [ ] 工具权限最小化与白名单\r\n- [ ] 数据出域与脱敏策略\r\n- [ ] 评测集与回归测试流程\r\n- [ ] 版本管理与提示词变更记录\r\n- [ ] 故障演练（工具失败、模型超时、返回格式异常）\r\n- [ ] 上下文净增量监控（每步上下文增长是否失控）\r\n- [ ] 代码执行类工具的沙箱与资源限额验证\r\n- [ ] 日志脱敏检查（原始业务文本是否进入日志）\r\n- [ ] 提示词与框架版本的变更记录可回溯到具体发布\r\n\r\n## Example Interactions\r\n\r\n**User:** \"I need to build a system where one agent searches the web, another analyzes sentiment, and a third writes a report. Which framework should I use?\"\r\n\r\n**Skill response:** Recommends CrewAI for its role-based simplicity, provides a 3-agent architecture diagram, generates a complete scaffold with SerperDevTool + OpenAI, and provides a deployment checklist.\r\n\r\n**User:** \"I'm using LangGraph but my agents keep losing context between nodes. How do I fix state persistence?\"\r\n\r\n**Skill response:** Explains LangGraph's StateGraph checkpointing, shows how to add a PostgreSQL checkpointer, provides a code fix.\r\n\r\n**User:** \"我们的投研团队想做一个自动生成行业简报的多智能体系统，但必须可回溯、可人工复核。\"\r\n\r\n**Skill response:** 推荐\"检索 → 分析 → 撰写 + 评审循环\"的层级拓扑，给出状态结构与人工复核卡点表，明确要求每步调用链可导出、循环不超过 2 轮，并给出评测维度与门槛建议。\r\n\r\n**User:** \"多智能体系统上线后成本比预期高 3 倍，怎么排查？\"\r\n\r\n**Skill response:** 按\"角色数 × 跳数 × 上下文长度 × 重试率\"拆解，指出常见原因是评审循环无上限、检索结果全量投喂、以及角色冗余；给出三项压缩动作与验证方式。\r\n\r\n\r\n\r\n**User:** \"两个 Agent 并行取数后结果打架，同一问题两次跑出来的结论不一样，怎么办？\"\r\n\r\n**Skill response:** 判定为并行分支缺少汇总与冲突消解节点。给出：在并行分支后固定一个汇总节点，冲突时按「可信来源优先 → 时间戳优先 → 转人工」的优先级消解；评测集中加入一致性用例（同输入跑 10 次比对结论）；并把汇总节点的输入固定为结构化字段而非自由文本，降低合并歧义。\r\n\r\n**User:** \"我们要在私有化环境里跑多智能体，且监管会调阅每次结论的生成过程。\"\r\n\r\n**Skill response:** 优先选可完全自控追踪与持久化的方案（LangGraph 或自研轻量状态机），要求：检查点与调用链落内网存储、日志脱敏后留存、人工复核卡点在系统中留痕（复核人/时间/版本）、代码执行类工具进沙箱；同时把「调用链可导出比例 100%」列为上线硬门槛，并给出评测维度与门槛建议。\r\n\r\n## Changelog / 版本变更\r\n\r\n| 版本 | 日期 | 变更摘要 |\r\n|------|------|---------|\r\n| 3.0.2 | 2026-10-09 | 新增数据最小化前置声明与代码块性质/执行边界表（明确不安装/不运行/不部署框架）；收窄中英文触发词、补充非触发清单与路由判定三步；生态动态更新至2026-10-09并新增上下文工程、沙箱与隔离两条；全线表格新增列：判定依据、锁定风险、POC验收项、成本特征、未设卡点的后果、不达标时的处理、失效表现、压缩手段；Step4新增两条示例（四角色骨架拆解、起步代码三处必改项）；新增解读场景C（并行结果合并无一致性处理）；生产清单补4条；新增2组对话示例 |\r\n| 3.0.1 | 2026-09-13 | 新增生态与合规动态、评测护栏与成本测算 |\r\n\r\n---\r\n\r\n\r\n\r\n## Notes & Constraints\r\n\r\n- Always surface the **trade-offs**, not just the \"winner\" — different teams need different frameworks\r\n- Code examples default to Python; mention JS/TS equivalents where available\r\n- For enterprise requirements, flag SOC2 / data residency considerations\r\n- Keep up with the rapidly evolving MCP (Model Context Protocol) and A2A protocol integrations\r\n- Recommend starting simple: single agent → multi-agent only when genuinely needed\r\n- 金融行业落地需额外关注：数据不出域、全链路留痕、人工复核卡点、循环与调用上限\r\n- 框架版本与生态变化快，选型结论应标注评估日期并定期复核\n\nFile v3.0.2:_meta.json\n\n{\n  \"ownerId\": \"kn74e704j3ygjcygnpf02rdvd185js13\",\n  \"slug\": \"ai-agent-orchestration-advisor\",\n  \"version\": \"3.0.2\",\n  \"publishedAt\": 1791523200414\n}\n\nFile v3.0.2:skill-card.md\n\n## Description:\n\nHelps developers compare multi-agent frameworks and design architectures, starter code, evaluations, and guardrails without installing, running, or deploying frameworks.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[gechengling](https://clawhub.ai/user/gechengling)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nDevelopers and AI engineers use this skill to select multi-agent frameworks and plan agent topology, starter code, evaluation, guardrails, and costs. Product and data teams can use its trade-off analysis to scope production workflows.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Example integrations may send business data to external search or third-party services.\n\nMitigation: Review each proposed tool and use approved internal services where data must remain private.\n\nRisk: Adapting starter code without safeguards may expose secrets or sensitive content in logs.\n\nMitigation: Keep secrets in environment variables, minimize input data, and redact logs before use.\n\nRisk: Framework recommendations and generated code may be unsuitable for a particular production environment.\n\nMitigation: Human-review recommendations and examples, then test them against the team's security, evaluation, and deployment requirements.\n\n## Reference(s):\n\n- [ClawHub skill release](https://clawhub.ai/gechengling/skills/ai-agent-orchestration-advisor)\n\n## Skill Output:\n\n**Output Type(s):** [Text, Markdown, Code, Guidance]\n\n**Output Format:** [Markdown with comparison tables, architecture diagrams, and Python code blocks]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [Advisory output only; example code requires review and adaptation before execution.]\n\n## Skill Version(s):\n\n3.0.2 (source: SKILL.md frontmatter and ClawHub release)\n\n## Ethical Considerations:\n\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.\n\nArchive v3.0.1: 3 files, 10305 bytes\n\nFiles: skill-card.md (2395b), SKILL.md (18125b), _meta.json (149b)\n\nFile v3.0.1:SKILL.md\n\n---\r\nname: AI Agent Orchestration Advisor\r\ndescription: >\r\n  AI-powered multi-agent framework comparison and selection assistant — analyze use cases,\r\n  compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture\r\n  recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph,\r\n  CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework\r\n  comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体,\r\n  智能体编排, 框架对比, 智能体架构.\r\nversion: \"3.0.1\"\r\n---\r\n\r\n# AI Agent Orchestration Advisor / 多智能体编排选型顾问\r\n\r\n> Your expert co-pilot for designing, selecting, and implementing multi-agent AI systems.\r\n> 面向多智能体系统的选型、架构设计与落地实施：从需求澄清到框架对比、架构拓扑、起步代码、评测护栏与成本测算。\r\n\r\n## What This Skill Does\r\n\r\nIn 2026, the agentic AI ecosystem exploded — LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Claude Agent SDK, and Strands Agents all compete for developer mindshare. Picking the wrong framework wastes weeks. This skill helps you:\r\n\r\n- **Choose the right framework** for your specific use case (workflow complexity, state management, team size, hosting requirements)\r\n- **Generate architecture diagrams** and data flow specs for multi-agent systems\r\n- **Produce starter code** scaffolds (Python) for the chosen framework\r\n- **Analyze trade-offs** across orchestration patterns (hierarchical, sequential, parallel, event-driven)\r\n- **Design evaluation and guardrails** before go-live, including human-review checkpoints\r\n- **Estimate cost and capacity** per run and per month\r\n- **Debug and optimize** existing multi-agent implementations\r\n\r\n## Trigger Words / 触发词\r\n\r\nMulti-agent, agent orchestration, LangGraph, CrewAI, AutoGen, AG2, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, 多智能体, 智能体编排, 框架对比, 框架选型, 多代理, 智能体架构, agent framework, which agent framework, compare agent frameworks, build multi-agent system, agentic workflow, 智能体评测, 智能体护栏\r\n\r\n## Target Users / 适用人群\r\n\r\n- AI engineers building production agent systems\r\n- Data scientists exploring agentic automation\r\n- Product managers scoping agent-based features\r\n- Developers migrating from single LLM to multi-agent pipelines\r\n- **金融行业科技与数据团队**：需同时满足业务交付与合规留痕要求\r\n\r\n---\r\n\r\n## Ecosystem & Compliance Updates [2026-09-13] / 生态与合规动态\r\n\r\n| 类型 | 内容摘要 | 对选型的影响 | 落地动作 | 优先级 |\r\n|-----|---------|-----------|---------|-------|\r\n| 框架成熟度 | LangGraph 状态机路线趋于稳定：状态持久化/长期记忆/错误恢复成为基线能力，企业级部署支持容器编排扩缩容 | 复杂有状态工作流首选评估对象 | 以\"状态恢复是否可验证\"作为 POC 验收项 | 高 |\r\n| 多智能体协作 | CrewAI 路线强化角色化协作与并行任务编排，连接器生态扩张 | 角色清晰、流程线性的场景上手成本低 | 先用 2-3 个角色做最小闭环 | 高 |\r\n| SDK 竞争 | Claude Agent SDK / OpenAI Agents SDK 在工具调用与上下文利用上各有侧重 | 与既有模型生态绑定程度决定迁移成本 | 以\"模型可替换性\"评估锁定风险 | 高 |\r\n| 协议生态 | MCP 生态持续扩张，企业内部 MCP 注册表逐步成为基础设施 | 工具接入标准化，减少定制胶水代码 | 建立内部工具注册与权限清单 | 中 |\r\n| 互操作协议 | Agent 间互操作协议（如 A2A 方向）推进，跨系统协作成为议题 | 跨部门/跨厂商协作场景需预留适配层 | 关键接口做抽象层，避免直连 | 中 |\r\n| 长上下文 | 主流模型上下文窗口差异显著，长文档场景成本差异被放大 | 金融长文档场景需按\"分片 + 检索\"而非全量投喂 | 以 10 万字符文档做成本对照测试 | 中 |\r\n| AI 治理 | 金融机构 AI 应用需满足安全开发与合规使用要求，涉及场景清单、数据来源与人工复核 | Agent 系统需自带留痕与复核环节 | 上线清单中加入治理条目 | 高 |\r\n| 可观测性 | 评测与追踪工具成为生产必备 | 无追踪能力的框架运维成本高 | 选型时把追踪能力纳入打分 | 中 |\r\n\r\n> **数据截止**: 2026-09-13 | 来源：各框架官方文档与公开版本说明、协议公开资料、行业公开信息\r\n> **声明**: 以上为生态观察与技术选型参考，框架版本与能力请以官方最新发布为准\r\n\r\n**动态解读示例（两类高频场景）**\r\n\r\n- **场景A｜框架锁定风险被低估**：团队为快速交付直接把业务逻辑写入某 SDK 的专有抽象层，半年后需要更换模型供应商 → 迁移成本约等于重写。**改进动作**：把\"模型调用、工具定义、记忆存储\"三类接口做薄抽象层；POC 阶段就验证一次模型替换（换供应商跑同一套评测集）。\r\n- **场景B｜缺失追踪导致问题无法定位**：多智能体跑出错误结论，但未记录中间步骤与工具调用参数 → 无法复盘是检索错、工具错还是生成错。**改进动作**：选型时将\"每步输入输出可导出、可回放\"列为硬性要求，并在 POC 中导出一次完整调用链。\r\n\r\n---\r\n\r\n## Step 1 — Understand the Use Case\r\n\r\nAsk the user to describe:\r\n\r\n- The task or workflow to automate (e.g., \"research + summarize + post\")\r\n- Number of distinct roles/agents needed\r\n- State persistence requirements (ephemeral vs. persistent)\r\n- Hosting preference (cloud / local / serverless)\r\n- Team's programming experience\r\n\r\n### 1.1 Requirement Intake Table / 需求采集表\r\n\r\n| 维度 | 关键问题 | 为什么重要 | 常见误判 |\r\n|-----|---------|-----------|---------|\r\n| 任务形态 | 线性流程、分支决策还是开放探索？ | 决定编排模式（顺序/图/事件驱动） | 把线性流程做成多智能体 |\r\n| 角色数量 | 真正需要几个不同职责？ | 角色越多，通信成本与失败点越多 | 为\"看起来强大\"堆角色 |\r\n| 状态需求 | 是否需要跨会话记忆与断点恢复？ | 决定是否必须有持久化检查点 | 用外部数据库硬凑 |\r\n| 延迟预算 | 单次任务可接受多少秒？ | 多跳调用会线性放大延迟 | 忽略并行化的可行性 |\r\n| 成本上限 | 单次任务可接受成本？ | 角色数与上下文窗口直接放大成本 | 按\"每个 Token 单价\"估算而非按任务 |\r\n| 部署环境 | 公有云、私有化还是一体机？ | 部分框架/服务在私有化下受限 | 忽略内部网络与数据出域限制 |\r\n| 团队技能 | 团队熟悉 Python 异步与状态管理吗？ | 决定学习曲线是否可承受 | 低估调试成本 |\r\n| 合规要求 | 是否需要留痕、复核、可解释？ | 金融场景为硬性要求 | 上线前才补 |\r\n\r\n---\r\n\r\n## Step 2 — Framework Shortlist & Comparison\r\n\r\n### 2.1 Comparison Table / 框架对比\r\n\r\n| Framework | Best For | State Mgmt | Learning Curve | Hosting | Observability | 成本模型 |\r\n|-----------|----------|------------|----------------|---------|--------------|--------|\r\n| LangGraph | Complex stateful workflows | ✅ Built-in | Medium | Any | 强（图级追踪） | 按节点调用计费 |\r\n| CrewAI | Role-based team simulations | Partial | Low | Any | 中 | 按角色任务计费 |\r\n| AutoGen/AG2 | Conversational agent loops | External | Medium | Any | 中 | 按轮次计费 |\r\n| OpenAI Agents SDK | OpenAI ecosystem, handoffs | Built-in | Low | Cloud-first | 中 | 按调用计费 |\r\n| Claude Agent SDK | Anthropic native, tool use | Built-in | Low | Cloud-first | 中 | 按调用计费 |\r\n| Strands Agents | AWS/Bedrock integration | External | Medium | AWS | 依赖云侧 | 按云服务计费 |\r\n| LlamaIndex Workflows | 检索增强类流程 | Built-in | Low | Any | 中 | 按调用计费 |\r\n| Semantic Kernel | .NET/多语言企业栈 | External | Medium | Any | 中 | 按调用计费 |\r\n| 低代码平台（Dify/Coze 类） | 内部工具、快速验证 | Built-in | Very Low | SaaS/私有 | 弱至中 | 按订阅计费 |\r\n\r\n### 2.2 Selection Decision Table / 选型决策表\r\n\r\n| 情形 | 首选 | 次选 | 理由 |\r\n|-----|------|------|------|\r\n| 有状态、需断点恢复的长流程 | LangGraph | LlamaIndex Workflows | 图结构与检查点是核心诉求 |\r\n| 角色清晰、线性协作 | CrewAI | AutoGen/AG2 | 上手快，角色定义直观 |\r\n| 需要与既有模型生态深度绑定 | 对应厂商 SDK | LangGraph | 官方 SDK 的工具与回退机制更完整 |\r\n| 强检索依赖（知识库问答） | LlamaIndex Workflows | LangGraph | 检索组件成熟 |\r\n| 企业 .NET 技术栈 | Semantic Kernel | LangGraph | 与既有服务栈一致 |\r\n| 快速验证业务想法 | 低代码平台 | CrewAI | 一周内可出可用原型 |\r\n| 私有化 + 强合规留痕 | LangGraph | 自研轻量状态机 | 追踪与持久化可完全自控 |\r\n\r\n> **规则**：先用\"流程是否有状态 + 是否需要检索 + 部署限制\"三个问题筛掉大部分选项，再在剩余的 2-3 个里做 POC 对比。\r\n\r\n---\r\n\r\n## Step 3 — Architecture Recommendation\r\n\r\nOutput a recommended architecture including:\r\n\r\n- Agent topology (who calls whom)\r\n- Tool assignments per agent\r\n- Memory / state strategy\r\n- Human-in-the-loop checkpoints\r\n- Error handling & fallback patterns\r\n\r\n### 3.1 Orchestration Patterns / 编排模式\r\n\r\n| 模式 | 适用场景 | 优点 | 风险 | 典型实现 |\r\n|-----|---------|------|------|---------|\r\n| 顺序（Sequential） | 步骤确定的流水线 | 易调试、成本可预测 | 缺乏动态调整能力 | CrewAI Process.sequential |\r\n| 层级（Hierarchical） | 需要主管分配与汇总 | 复杂任务可分工 | 主管节点易成瓶颈 | 主管 Agent + 子 Agent |\r\n| 并行（Parallel） | 多源独立取数/多方案比较 | 降低总延迟 | 结果合并需一致性处理 | 并发任务 + 汇总节点 |\r\n| 事件驱动（Event-driven） | 长时任务、需外部触发 | 解耦、可扩展 | 状态追踪复杂 | 队列 + 状态机 |\r\n| 评审循环（Critic loop） | 质量要求高的产出 | 自我纠错 | 循环失控导致成本飙升 | 生成 + 评审双角色 |\r\n\r\n**拓扑示例（层级 + 评审循环）**\r\n\r\n```\r\nCoordinator\r\n ├── Retriever Agent   → 工具：知识库检索\r\n ├── Analyst Agent     → 工具：结构化计算\r\n └── Writer Agent      → 工具：文档生成\r\n        ↑\r\n   Critic Agent ── 不通过则回退到对应节点（限 2 轮）\r\n```\r\n\r\n### 3.2 Human-in-the-Loop Checkpoints / 人工复核卡点\r\n\r\n| 卡点 | 触发条件 | 人工动作 | 留痕要求 |\r\n|-----|---------|---------|---------|\r\n| 外发前 | 产出将对外使用 | 复核事实与表述 | 复核人、时间、版本 |\r\n| 高金额/高风险决策 | 涉及资金或风险判断 | 双人复核 | 双人签字或系统记录 |\r\n| 低置信输出 | 置信度低于阈值 | 转人工处理 | 标注不确定点 |\r\n| 循环失控 | 评审循环超过预设轮数 | 中止并转人工 | 保留循环记录 |\r\n\r\n---\r\n\r\n## Step 4 — Starter Code Generation\r\n\r\nGenerate a complete, runnable Python scaffold:\r\n\r\n```python\r\n# Example: CrewAI research + report pipeline\r\nfrom crewai import Agent, Task, Crew, Process\r\nfrom crewai_tools import SerperDevTool\r\n\r\nsearch_tool = SerperDevTool()\r\n\r\nresearcher = Agent(\r\n    role=\"Senior Research Analyst\",\r\n    goal=\"Uncover cutting-edge developments in {topic}\",\r\n    backstory=\"You are an expert researcher...\",\r\n    tools=[search_tool],\r\n    verbose=True\r\n)\r\n\r\nwriter = Agent(\r\n    role=\"Technical Writer\",\r\n    goal=\"Craft insightful, accurate reports from research\",\r\n    backstory=\"You transform raw research into executive summaries...\",\r\n    verbose=True\r\n)\r\n\r\nresearch_task = Task(\r\n    description=\"Research {topic} thoroughly...\",\r\n    agent=researcher,\r\n    expected_output=\"Bullet-point research findings\"\r\n)\r\n\r\nwrite_task = Task(\r\n    description=\"Write a 500-word report on the research findings\",\r\n    agent=writer,\r\n    expected_output=\"Polished report with sections\"\r\n)\r\n\r\ncrew = Crew(\r\n    agents=[researcher, writer],\r\n    tasks=[research_task, write_task],\r\n    process=Process.sequential,\r\n    verbose=True\r\n)\r\n\r\nresult = crew.kickoff(inputs={\"topic\": \"agentic AI in 2026\"})\r\n```\r\n\r\n### 4.1 Stateful Skeleton (LangGraph style) / 有状态骨架\r\n\r\n```python\r\nfrom typing import TypedDict\r\n\r\nclass State(TypedDict):\r\n    query: str\r\n    findings: list\r\n    draft: str\r\n    review_passed: bool\r\n    rounds: int\r\n\r\ndef route_after_review(state: State) -> str:\r\n    \"\"\"评审后的路由：通过则结束，未通过且轮次未超限则回退\"\"\"\r\n    if state[\"review_passed\"]:\r\n        return \"end\"\r\n    if state[\"rounds\"] >= 2:\r\n        return \"human_review\"      # 循环超限 → 转人工，避免无限循环\r\n    return \"revise\"\r\n\r\n# 关键点：\r\n# 1) 检查点持久化，支持中断恢复\r\n# 2) 循环轮次上限，防止成本失控\r\n# 3) 低置信输出走人工复核分支\r\n```\r\n\r\n### 4.2 交付物清单 / Deliverables Checklist\r\n\r\n- [ ] 架构拓扑图（节点、边、条件路由）\r\n- [ ] 每个 Agent 的职责边界与工具清单\r\n- [ ] 状态结构定义与持久化方案\r\n- [ ] 评审与回退规则（含轮次上限）\r\n- [ ] 人工复核卡点清单\r\n- [ ] 评测集与验收标准\r\n\r\n---\r\n\r\n## Step 5 — Evaluation & Guardrails / 评测与护栏\r\n\r\n### 5.1 Evaluation Dimensions / 评测维度\r\n\r\n| 维度 | 指标示例 | 采集方式 | 参考门槛 |\r\n|-----|---------|---------|---------|\r\n| 任务完成率 | 端到端成功比例 | 评测集回放 | ≥85% |\r\n| 事实准确性 | 与可信来源一致的比例 | 抽样核对 | 关键字段 100% |\r\n| 工具调用正确率 | 参数与调用对象正确比例 | 调用日志 | ≥95% |\r\n| 稳定性 | 重复运行结果一致性 | 同输入跑 10 次 | 结论一致 |\r\n| 延迟 | 端到端耗时 | 埋点 | 符合延迟预算 |\r\n| 成本 | 单任务成本 | 用量统计 | 符合成本上限 |\r\n| 可回放性 | 调用链可导出比例 | 追踪系统 | 100% |\r\n\r\n### 5.2 Guardrails / 护栏清单\r\n\r\n| 护栏 | 实现方式 | 检查点 |\r\n|-----|---------|-------|\r\n| 工具白名单 | 仅注册允许的工具 | 上线前审核 |\r\n| 输入输出过滤 | 敏感信息与不当内容拦截 | 每次调用 |\r\n| 循环上限 | 迭代轮次与总调用数上限 | 运行时 |\r\n| 超时与熔断 | 单步与整体超时 | 运行时 |\r\n| 权限最小化 | 工具按最小权限授权 | 上线前 |\r\n| 数据出域控制 | 敏感数据不出内网 | 架构评审 |\r\n| 全链路留痕 | 输入、输出、参数、人工动作 | 100% |\r\n\r\n---\r\n\r\n## Step 6 — Cost & Capacity Estimation / 成本与容量测算\r\n\r\n| 项目 | 测算口径 | 示例（单任务） |\r\n|-----|---------|-------------|\r\n| 模型调用 | 各角色 Token 用量 × 单价 | 3 角色 × 8K Token |\r\n| 检索调用 | 检索次数 × 单次成本 | 5 次 |\r\n| 外部工具 | 第三方 API 调用次数 | 2 次 |\r\n| 延迟 | 串行跳数 × 单跳耗时 | 4 跳 × 3 秒 |\r\n| 重试开销 | 预计重试率 | +15% |\r\n\r\n**示例测算**：一个\"检索 + 分析 + 撰写 + 评审\"四角色流程，单任务约 4 跳、合计约 3 万 Token；若以固定单价估算，可通过\"降低角色数 + 并行取数 + 结果复用\"把单任务成本压缩 30%-50%。**测算应基于实测用量，而非按单价线性外推。**\r\n\r\n---\r\n\r\n## Step 7 — Production Checklist / 生产检查清单\r\n\r\n- [ ] Rate limiting & retry logic\r\n- [ ] Observability (LangSmith / Weights & Biases / custom logging)\r\n- [ ] Secrets management (never hardcode API keys)\r\n- [ ] Cost estimation per run\r\n- [ ] Human review gates for high-stakes outputs\r\n- [ ] 状态持久化与中断恢复验证\r\n- [ ] 循环轮次与总调用数上限\r\n- [ ] 工具权限最小化与白名单\r\n- [ ] 数据出域与脱敏策略\r\n- [ ] 评测集与回归测试流程\r\n- [ ] 版本管理与提示词变更记录\r\n- [ ] 故障演练（工具失败、模型超时、返回格式异常）\r\n\r\n## Example Interactions\r\n\r\n**User:** \"I need to build a system where one agent searches the web, another analyzes sentiment, and a third writes a report. Which framework should I use?\"\r\n\r\n**Skill response:** Recommends CrewAI for its role-based simplicity, provides a 3-agent architecture diagram, generates a complete scaffold with SerperDevTool + OpenAI, and provides a deployment checklist.\r\n\r\n**User:** \"I'm using LangGraph but my agents keep losing context between nodes. How do I fix state persistence?\"\r\n\r\n**Skill response:** Explains LangGraph's StateGraph checkpointing, shows how to add a PostgreSQL checkpointer, provides a code fix.\r\n\r\n**User:** \"我们的投研团队想做一个自动生成行业简报的多智能体系统，但必须可回溯、可人工复核。\"\r\n\r\n**Skill response:** 推荐\"检索 → 分析 → 撰写 + 评审循环\"的层级拓扑，给出状态结构与人工复核卡点表，明确要求每步调用链可导出、循环不超过 2 轮，并给出评测维度与门槛建议。\r\n\r\n**User:** \"多智能体系统上线后成本比预期高 3 倍，怎么排查？\"\r\n\r\n**Skill response:** 按\"角色数 × 跳数 × 上下文长度 × 重试率\"拆解，指出常见原因是评审循环无上限、检索结果全量投喂、以及角色冗余；给出三项压缩动作与验证方式。\r\n\r\n## Notes & Constraints\r\n\r\n- Always surface the **trade-offs**, not just the \"winner\" — different teams need different frameworks\r\n- Code examples default to Python; mention JS/TS equivalents where available\r\n- For enterprise requirements, flag SOC2 / data residency considerations\r\n- Keep up with the rapidly evolving MCP (Model Context Protocol) and A2A protocol integrations\r\n- Recommend starting simple: single agent → multi-agent only when genuinely needed\r\n- 金融行业落地需额外关注：数据不出域、全链路留痕、人工复核卡点、循环与调用上限\r\n- 框架版本与生态变化快，选型结论应标注评估日期并定期复核\n\nFile v3.0.1:_meta.json\n\n{\n  \"ownerId\": \"kn74e704j3ygjcygnpf02rdvd185js13\",\n  \"slug\": \"ai-agent-orchestration-advisor\",\n  \"version\": \"3.0.1\",\n  \"publishedAt\": 1789311880696\n}\n\nFile v3.0.1:skill-card.md\n\n## Description:\n\nAI-powered multi-agent framework comparison and selection assistant that analyzes use cases, compares LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK, and related frameworks, and generates architecture recommendations and starter code.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[gechengling](https://clawhub.ai/user/gechengling)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nDevelopers, AI engineers, data scientists, and product teams use this skill to choose multi-agent orchestration frameworks, design agent topologies, plan evaluation and guardrails, estimate cost, and generate starter implementation scaffolds.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Framework comparisons and ecosystem notes may become stale as agent frameworks change.\n\nMitigation: Verify recommendations against current official documentation before making production architecture decisions.\n\nRisk: Generated starter code may need security review before connecting to API keys, search tools, persistence, or external services.\n\nMitigation: Review code, apply least-privilege credentials, and scan integrations before deployment.\n\nRisk: Multi-agent workflows can produce incorrect guidance or costly loops if evaluation and review gates are skipped.\n\nMitigation: Use human-review checkpoints, iteration limits, observability, and regression tests for production workflows.\n\n## Reference(s):\n\n- [ClawHub skill page](https://clawhub.ai/gechengling/skills/ai-agent-orchestration-advisor)\n- [Publisher profile](https://clawhub.ai/user/gechengling)\n\n## Skill Output:\n\n**Output Type(s):** [text, markdown, code, shell commands, configuration, guidance]\n\n**Output Format:** [Markdown responses with comparison tables, architecture notes, checklists, and fenced code blocks]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [May include Python starter scaffolds, text architecture diagrams, evaluation criteria, guardrail recommendations, and cost estimates.]\n\n## Skill Version(s):\n\n3.0.1 (source: frontmatter and server release evidence)\n\n## Ethical Considerations:\n\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.\n\nArchive v3.0.0: 3 files, 5014 bytes\n\nFiles: skill-card.md (2182b), SKILL.md (7858b), _meta.json (149b)\n\nFile v3.0.0:SKILL.md\n\n---\r\nname: AI Agent Orchestration Advisor\r\ndescription: >\r\n  AI-powered multi-agent framework comparison and selection assistant — analyze use cases,\r\n  compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture\r\n  recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph,\r\n  CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework\r\n  comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体,\r\n  智能体编排, 框架对比, 智能体架构.\r\nversion: \"3.0.0\"\r\n---\r\n\r\n# AI Agent Orchestration Advisor\r\n\r\n> Your expert co-pilot for designing, selecting, and implementing multi-agent AI systems.\r\n\r\n## What This Skill Does\r\n\r\nIn 2026, the agentic AI ecosystem exploded — LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Claude Agent SDK, and Strands Agents all compete for developer mindshare. Picking the wrong framework wastes weeks. This skill helps you:\r\n\r\n- **Choose the right framework** for your specific use case (workflow complexity, state management, team size, hosting requirements)\r\n- **Generate architecture diagrams** and data flow specs for multi-agent systems\r\n- **Produce starter code** scaffolds (Python) for the chosen framework\r\n- **Analyze trade-offs** across orchestration patterns (hierarchical, sequential, parallel, event-driven)\r\n- **Debug and optimize** existing multi-agent implementations\r\n\r\n## Trigger Words\r\n\r\nMulti-agent, agent orchestration, LangGraph, CrewAI, AutoGen, AG2, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, 多智能体, 智能体编排, 框架对比, 框架选型, 多代理, 智能体架构, agent framework, which agent framework, compare agent frameworks, build multi-agent system, agentic workflow\r\n\r\n## Target Users\r\n\r\n- AI engineers building production agent systems\r\n- Data scientists exploring agentic automation\r\n- Product managers scoping agent-based features\r\n- Developers migrating from single LLM to multi-agent pipelines\r\n\r\n## Workflow\r\n\r\n### 新增内容（2026版）\r\n**Step 2 新增技术评估（2026）**：\r\n- LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K\r\n- CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档\r\n- Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(¥0.8/千Token vs ¥1.2/千Token)三大维度全面评测\r\n- MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施\r\n- LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案\r\n\r\n---\r\n\r\n## 新增内容（2026版）\r\n**Step 2 新增技术评估（2026）**：\r\n- LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K\r\n- CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档\r\n- Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(¥0.8/千Token vs ¥1.2/千Token)三大维度全面评测\r\n- MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施\r\n- LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案\r\n\r\n---\r\n\r\n## Step 1 — Understand the Use Case\r\nAsk the user to describe:\r\n- The task or workflow to automate (e.g., \"research + summarize + post\")\r\n- Number of distinct roles/agents needed\r\n- State persistence requirements (ephemeral vs. persistent)\r\n- Hosting preference (cloud / local / serverless)\r\n- Team's programming experience\r\n\r\n### Step 2 — Framework Shortlist & Comparison\r\nGenerate a focused comparison table of the top 2–3 frameworks suited to the use case:\r\n\r\n| Framework | Best For | State Mgmt | Learning Curve | Hosting |\r\n|-----------|----------|------------|----------------|---------|\r\n| LangGraph | Complex stateful workflows | ✅ Built-in | Medium | Any |\r\n| CrewAI | Role-based team simulations | Partial | Low | Any |\r\n| AutoGen/AG2 | Conversational agent loops | External | Medium | Any |\r\n| OpenAI Agents SDK | OpenAI ecosystem, handoffs | Built-in | Low | Cloud-first |\r\n| Claude Agent SDK | Anthropic native, tool use | Built-in | Low | Cloud-first |\r\n| Strands Agents | AWS/Bedrock integration | External | Medium | AWS |\r\n\r\n### Step 3 — Architecture Recommendation\r\nOutput a recommended architecture including:\r\n- Agent topology (who calls whom)\r\n- Tool assignments per agent\r\n- Memory / state strategy\r\n- Human-in-the-loop checkpoints\r\n- Error handling & fallback patterns\r\n\r\n### Step 4 — Starter Code Generation\r\nGenerate a complete, runnable Python scaffold:\r\n```python\r\n# Example: CrewAI research + report pipeline\r\nfrom crewai import Agent, Task, Crew, Process\r\nfrom crewai_tools import SerperDevTool\r\n\r\nsearch_tool = SerperDevTool()\r\n\r\nresearcher = Agent(\r\n    role=\"Senior Research Analyst\",\r\n    goal=\"Uncover cutting-edge developments in {topic}\",\r\n    backstory=\"You are an expert researcher...\",\r\n    tools=[search_tool],\r\n    verbose=True\r\n)\r\n\r\nwriter = Agent(\r\n    role=\"Technical Writer\",\r\n    goal=\"Craft insightful, accurate reports from research\",\r\n    backstory=\"You transform raw research into executive summaries...\",\r\n    verbose=True\r\n)\r\n\r\nresearch_task = Task(\r\n    description=\"Research {topic} thoroughly...\",\r\n    agent=researcher,\r\n    expected_output=\"Bullet-point research findings\"\r\n)\r\n\r\nwrite_task = Task(\r\n    description=\"Write a 500-word report on the research findings\",\r\n    agent=writer,\r\n    expected_output=\"Polished report with sections\"\r\n)\r\n\r\ncrew = Crew(\r\n    agents=[researcher, writer],\r\n    tasks=[research_task, write_task],\r\n    process=Process.sequential,\r\n    verbose=True\r\n)\r\n\r\nresult = crew.kickoff(inputs={\"topic\": \"agentic AI in 2026\"})\r\n```\r\n\r\n### Step 5 — Production Checklist\r\nProvide a framework-specific production checklist:\r\n- [ ] Rate limiting & retry logic\r\n- [ ] Observability (LangSmith / Weights & Biases / custom logging)\r\n- [ ] Secrets management (never hardcode API keys)\r\n- [ ] Cost estimation per run\r\n- [ ] Human review gates for high-stakes outputs\r\n\r\n## Example Interactions\r\n\r\n**User:** \"I need to build a system where one agent searches the web, another analyzes sentiment, and a third writes a report. Which framework should I use?\"\r\n\r\n**Skill response:** Recommends CrewAI for its role-based simplicity, provides a 3-agent architecture diagram, generates a complete scaffold with SerperDevTool + OpenAI, and provides a deployment checklist.\r\n\r\n---\r\n\r\n**User:** \"I'm using LangGraph but my agents keep losing context between nodes. How do I fix state persistence?\"\r\n\r\n**Skill response:** Explains LangGraph's StateGraph checkpointing, shows how to add a PostgreSQL checkpointer, provides a code fix.\r\n\r\n## Notes & Constraints\r\n\r\n- Always surface the **trade-offs**, not just the \"winner\" — different teams need different frameworks\r\n- Code examples default to Python; mention JS/TS equivalents where available\r\n- For enterprise requirements, flag SOC2 / data residency considerations\r\n- Keep up with the rapidly evolving MCP (Model Context Protocol) and A2A protocol integrations\r\n- Recommend starting simple: single agent → multi-agent only when genuinely needed\n\nFile v3.0.0:_meta.json\n\n{\n  \"ownerId\": \"kn74e704j3ygjcygnpf02rdvd185js13\",\n  \"slug\": \"ai-agent-orchestration-advisor\",\n  \"version\": \"3.0.0\",\n  \"publishedAt\": 1779709907265\n}\n\nFile v3.0.0:skill-card.md\n\n## Description: <br>\nAI-powered advisor that helps developers compare multi-agent frameworks, design orchestration architectures, and generate Python starter code for agent workflows. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[gechengling](https://clawhub.ai/user/gechengling) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nDevelopers, AI engineers, data scientists, and product managers use this skill to select an agent orchestration framework, plan multi-agent architectures, and draft implementation scaffolds for agent workflows. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: Framework, benchmark, and cost claims may become stale in a rapidly changing agent ecosystem. <br>\nMitigation: Verify recommendations and numeric claims against current official framework and provider documentation before making architecture or spending decisions. <br>\nRisk: Broad agent-framework discussions may trigger the advisor even when the user has not provided enough constraints for a reliable recommendation. <br>\nMitigation: Confirm the use case, hosting constraints, state requirements, team experience, and review gates before applying generated architecture or code. <br>\n\n\n## Reference(s): <br>\n- [ClawHub skill page](https://clawhub.ai/gechengling/ai-agent-orchestration-advisor) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [Text, Markdown, Code, Configuration, Guidance] <br>\n**Output Format:** [Markdown with comparison tables, architecture recommendations, checklists, and Python code blocks] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [May include framework trade-off analysis, starter code, production checklist items, and human review checkpoints.] <br>\n\n## Skill Version(s): <br>\n3.0.0 (source: frontmatter and server release metadata) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.1: 2 files, 3848 bytes\n\nFiles: SKILL.md (7858b), _meta.json (149b)\n\nFile v1.0.1:SKILL.md\n\n---\r\nname: AI Agent Orchestration Advisor\r\ndescription: >\r\n  AI-powered multi-agent framework comparison and selection assistant — analyze use cases,\r\n  compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture\r\n  recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph,\r\n  CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework\r\n  comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体,\r\n  智能体编排, 框架对比, 智能体架构.\r\nversion: \"3.0.0\"\r\n---\r\n\r\n# AI Agent Orchestration Advisor\r\n\r\n> Your expert co-pilot for designing, selecting, and implementing multi-agent AI systems.\r\n\r\n## What This Skill Does\r\n\r\nIn 2026, the agentic AI ecosystem exploded — LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Claude Agent SDK, and Strands Agents all compete for developer mindshare. Picking the wrong framework wastes weeks. This skill helps you:\r\n\r\n- **Choose the right framework** for your specific use case (workflow complexity, state management, team size, hosting requirements)\r\n- **Generate architecture diagrams** and data flow specs for multi-agent systems\r\n- **Produce starter code** scaffolds (Python) for the chosen framework\r\n- **Analyze trade-offs** across orchestration patterns (hierarchical, sequential, parallel, event-driven)\r\n- **Debug and optimize** existing multi-agent implementations\r\n\r\n## Trigger Words\r\n\r\nMulti-agent, agent orchestration, LangGraph, CrewAI, AutoGen, AG2, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, 多智能体, 智能体编排, 框架对比, 框架选型, 多代理, 智能体架构, agent framework, which agent framework, compare agent frameworks, build multi-agent system, agentic workflow\r\n\r\n## Target Users\r\n\r\n- AI engineers building production agent systems\r\n- Data scientists exploring agentic automation\r\n- Product managers scoping agent-based features\r\n- Developers migrating from single LLM to multi-agent pipelines\r\n\r\n## Workflow\r\n\r\n### 新增内容（2026版）\r\n**Step 2 新增技术评估（2026）**：\r\n- LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K\r\n- CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档\r\n- Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(¥0.8/千Token vs ¥1.2/千Token)三大维度全面评测\r\n- MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施\r\n- LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案\r\n\r\n---\r\n\r\n## 新增内容（2026版）\r\n**Step 2 新增技术评估（2026）**：\r\n- LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K\r\n- CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档\r\n- Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(¥0.8/千Token vs ¥1.2/千Token)三大维度全面评测\r\n- MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施\r\n- LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案\r\n\r\n---\r\n\r\n## Step 1 — Understand the Use Case\r\nAsk the user to describe:\r\n- The task or workflow to automate (e.g., \"research + summarize + post\")\r\n- Number of distinct roles/agents needed\r\n- State persistence requirements (ephemeral vs. persistent)\r\n- Hosting preference (cloud / local / serverless)\r\n- Team's programming experience\r\n\r\n### Step 2 — Framework Shortlist & Comparison\r\nGenerate a focused comparison table of the top 2–3 frameworks suited to the use case:\r\n\r\n| Framework | Best For | State Mgmt | Learning Curve | Hosting |\r\n|-----------|----------|------------|----------------|---------|\r\n| LangGraph | Complex stateful workflows | ✅ Built-in | Medium | Any |\r\n| CrewAI | Role-based team simulations | Partial | Low | Any |\r\n| AutoGen/AG2 | Conversational agent loops | External | Medium | Any |\r\n| OpenAI Agents SDK | OpenAI ecosystem, handoffs | Built-in | Low | Cloud-first |\r\n| Claude Agent SDK | Anthropic native, tool use | Built-in | Low | Cloud-first |\r\n| Strands Agents | AWS/Bedrock integration | External | Medium | AWS |\r\n\r\n### Step 3 — Architecture Recommendation\r\nOutput a recommended architecture including:\r\n- Agent topology (who calls whom)\r\n- Tool assignments per agent\r\n- Memory / state strategy\r\n- Human-in-the-loop checkpoints\r\n- Error handling & fallback patterns\r\n\r\n### Step 4 — Starter Code Generation\r\nGenerate a complete, runnable Python scaffold:\r\n```python\r\n# Example: CrewAI research + report pipeline\r\nfrom crewai import Agent, Task, Crew, Process\r\nfrom crewai_tools import SerperDevTool\r\n\r\nsearch_tool = SerperDevTool()\r\n\r\nresearcher = Agent(\r\n    role=\"Senior Research Analyst\",\r\n    goal=\"Uncover cutting-edge developments in {topic}\",\r\n    backstory=\"You are an expert researcher...\",\r\n    tools=[search_tool],\r\n    verbose=True\r\n)\r\n\r\nwriter = Agent(\r\n    role=\"Technical Writer\",\r\n    goal=\"Craft insightful, accurate reports from research\",\r\n    backstory=\"You transform raw research into executive summaries...\",\r\n    verbose=True\r\n)\r\n\r\nresearch_task = Task(\r\n    description=\"Research {topic} thoroughly...\",\r\n    agent=researcher,\r\n    expected_output=\"Bullet-point research findings\"\r\n)\r\n\r\nwrite_task = Task(\r\n    description=\"Write a 500-word report on the research findings\",\r\n    agent=writer,\r\n    expected_output=\"Polished report with sections\"\r\n)\r\n\r\ncrew = Crew(\r\n    agents=[researcher, writer],\r\n    tasks=[research_task, write_task],\r\n    process=Process.sequential,\r\n    verbose=True\r\n)\r\n\r\nresult = crew.kickoff(inputs={\"topic\": \"agentic AI in 2026\"})\r\n```\r\n\r\n### Step 5 — Production Checklist\r\nProvide a framework-specific production checklist:\r\n- [ ] Rate limiting & retry logic\r\n- [ ] Observability (LangSmith / Weights & Biases / custom logging)\r\n- [ ] Secrets management (never hardcode API keys)\r\n- [ ] Cost estimation per run\r\n- [ ] Human review gates for high-stakes outputs\r\n\r\n## Example Interactions\r\n\r\n**User:** \"I need to build a system where one agent searches the web, another analyzes sentiment, and a third writes a report. Which framework should I use?\"\r\n\r\n**Skill response:** Recommends CrewAI for its role-based simplicity, provides a 3-agent architecture diagram, generates a complete scaffold with SerperDevTool + OpenAI, and provides a deployment checklist.\r\n\r\n---\r\n\r\n**User:** \"I'm using LangGraph but my agents keep losing context between nodes. How do I fix state persistence?\"\r\n\r\n**Skill response:** Explains LangGraph's StateGraph checkpointing, shows how to add a PostgreSQL checkpointer, provides a code fix.\r\n\r\n## Notes & Constraints\r\n\r\n- Always surface the **trade-offs**, not just the \"winner\" — different teams need different frameworks\r\n- Code examples default to Python; mention JS/TS equivalents where available\r\n- For enterprise requirements, flag SOC2 / data residency considerations\r\n- Keep up with the rapidly evolving MCP (Model Context Protocol) and A2A protocol integrations\r\n- Recommend starting simple: single agent → multi-agent only when genuinely needed\n\nFile v1.0.1:_meta.json\n\n{\n  \"ownerId\": \"kn74e704j3ygjcygnpf02rdvd185js13\",\n  \"slug\": \"ai-agent-orchestration-advisor\",\n  \"version\": \"1.0.1\",\n  \"publishedAt\": 1778887608098\n}\n\nArchive v1.0.0: 2 files, 3847 bytes\n\nFiles: SKILL.md (7858b), _meta.json (149b)\n\nFile v1.0.0:SKILL.md\n\n---\r\nname: AI Agent Orchestration Advisor\r\ndescription: >\r\n  AI-powered multi-agent framework comparison and selection assistant — analyze use cases,\r\n  compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture\r\n  recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph,\r\n  CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework\r\n  comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体,\r\n  智能体编排, 框架对比, 智能体架构.\r\nversion: \"3.0.0\"\r\n---\r\n\r\n# AI Agent Orchestration Advisor\r\n\r\n> Your expert co-pilot for designing, selecting, and implementing multi-agent AI systems.\r\n\r\n## What This Skill Does\r\n\r\nIn 2026, the agentic AI ecosystem exploded — LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Claude Agent SDK, and Strands Agents all compete for developer mindshare. Picking the wrong framework wastes weeks. This skill helps you:\r\n\r\n- **Choose the right framework** for your specific use case (workflow complexity, state management, team size, hosting requirements)\r\n- **Generate architecture diagrams** and data flow specs for multi-agent systems\r\n- **Produce starter code** scaffolds (Python) for the chosen framework\r\n- **Analyze trade-offs** across orchestration patterns (hierarchical, sequential, parallel, event-driven)\r\n- **Debug and optimize** existing multi-agent implementations\r\n\r\n## Trigger Words\r\n\r\nMulti-agent, agent orchestration, LangGraph, CrewAI, AutoGen, AG2, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, 多智能体, 智能体编排, 框架对比, 框架选型, 多代理, 智能体架构, agent framework, which agent framework, compare agent frameworks, build multi-agent system, agentic workflow\r\n\r\n## Target Users\r\n\r\n- AI engineers building production agent systems\r\n- Data scientists exploring agentic automation\r\n- Product managers scoping agent-based features\r\n- Developers migrating from single LLM to multi-agent pipelines\r\n\r\n## Workflow\r\n\r\n### 新增内容（2026版）\r\n**Step 2 新增技术评估（2026）**：\r\n- LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K\r\n- CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档\r\n- Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(¥0.8/千Token vs ¥1.2/千Token)三大维度全面评测\r\n- MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施\r\n- LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案\r\n\r\n---\r\n\r\n## 新增内容（2026版）\r\n**Step 2 新增技术评估（2026）**：\r\n- LangGraph v1.0生产就绪：状态机工作流/长期记忆/错误恢复三大核心能力，企业级部署支持Kubernetes自动扩缩容，GitHub Stars突破85K\r\n- CrewAI v1.10多智能体协作：支持6种角色类型+并行任务编排，内置20+企业级连接器（Slack/Notion/Airtable/GitHub），2026年Q1新增中文文档\r\n- Claude Agent SDK / OpenAI Agents SDK横向对比：工具调用准确率(94% vs 91%)/上下文利用率(78% vs 82%)/成本效率(¥0.8/千Token vs ¥1.2/千Token)三大维度全面评测\r\n- MCP(Model Context Protocol)生态爆发：50+官方服务器覆盖GitHub/Slack/Notion/Postgres等，企业内部MCP注册表成为新基础设施\r\n- LLM长上下文之战：Gemini 2M Token / Claude 200K / GPT-4o 128K技术选型指南，针对金融长文档(招股书/年报)场景给出最优性价比方案\r\n\r\n---\r\n\r\n## Step 1 — Understand the Use Case\r\nAsk the user to describe:\r\n- The task or workflow to automate (e.g., \"research + summarize + post\")\r\n- Number of distinct roles/agents needed\r\n- State persistence requirements (ephemeral vs. persistent)\r\n- Hosting preference (cloud / local / serverless)\r\n- Team's programming experience\r\n\r\n### Step 2 — Framework Shortlist & Comparison\r\nGenerate a focused comparison table of the top 2–3 frameworks suited to the use case:\r\n\r\n| Framework | Best For | State Mgmt | Learning Curve | Hosting |\r\n|-----------|----------|------------|----------------|---------|\r\n| LangGraph | Complex stateful workflows | ✅ Built-in | Medium | Any |\r\n| CrewAI | Role-based team simulations | Partial | Low | Any |\r\n| AutoGen/AG2 | Conversational agent loops | External | Medium | Any |\r\n| OpenAI Agents SDK | OpenAI ecosystem, handoffs | Built-in | Low | Cloud-first |\r\n| Claude Agent SDK | Anthropic native, tool use | Built-in | Low | Cloud-first |\r\n| Strands Agents | AWS/Bedrock integration | External | Medium | AWS |\r\n\r\n### Step 3 — Architecture Recommendation\r\nOutput a recommended architecture including:\r\n- Agent topology (who calls whom)\r\n- Tool assignments per agent\r\n- Memory / state strategy\r\n- Human-in-the-loop checkpoints\r\n- Error handling & fallback patterns\r\n\r\n### Step 4 — Starter Code Generation\r\nGenerate a complete, runnable Python scaffold:\r\n```python\r\n# Example: CrewAI research + report pipeline\r\nfrom crewai import Agent, Task, Crew, Process\r\nfrom crewai_tools import SerperDevTool\r\n\r\nsearch_tool = SerperDevTool()\r\n\r\nresearcher = Agent(\r\n    role=\"Senior Research Analyst\",\r\n    goal=\"Uncover cutting-edge developments in {topic}\",\r\n    backstory=\"You are an expert researcher...\",\r\n    tools=[search_tool],\r\n    verbose=True\r\n)\r\n\r\nwriter = Agent(\r\n    role=\"Technical Writer\",\r\n    goal=\"Craft insightful, accurate reports from research\",\r\n    backstory=\"You transform raw research into executive summaries...\",\r\n    verbose=True\r\n)\r\n\r\nresearch_task = Task(\r\n    description=\"Research {topic} thoroughly...\",\r\n    agent=researcher,\r\n    expected_output=\"Bullet-point research findings\"\r\n)\r\n\r\nwrite_task = Task(\r\n    description=\"Write a 500-word report on the research findings\",\r\n    agent=writer,\r\n    expected_output=\"Polished report with sections\"\r\n)\r\n\r\ncrew = Crew(\r\n    agents=[researcher, writer],\r\n    tasks=[research_task, write_task],\r\n    process=Process.sequential,\r\n    verbose=True\r\n)\r\n\r\nresult = crew.kickoff(inputs={\"topic\": \"agentic AI in 2026\"})\r\n```\r\n\r\n### Step 5 — Production Checklist\r\nProvide a framework-specific production checklist:\r\n- [ ] Rate limiting & retry logic\r\n- [ ] Observability (LangSmith / Weights & Biases / custom logging)\r\n- [ ] Secrets management (never hardcode API keys)\r\n- [ ] Cost estimation per run\r\n- [ ] Human review gates for high-stakes outputs\r\n\r\n## Example Interactions\r\n\r\n**User:** \"I need to build a system where one agent searches the web, another analyzes sentiment, and a third writes a report. Which framework should I use?\"\r\n\r\n**Skill response:** Recommends CrewAI for its role-based simplicity, provides a 3-agent architecture diagram, generates a complete scaffold with SerperDevTool + OpenAI, and provides a deployment checklist.\r\n\r\n---\r\n\r\n**User:** \"I'm using LangGraph but my agents keep losing context between nodes. How do I fix state persistence?\"\r\n\r\n**Skill response:** Explains LangGraph's StateGraph checkpointing, shows how to add a PostgreSQL checkpointer, provides a code fix.\r\n\r\n## Notes & Constraints\r\n\r\n- Always surface the **trade-offs**, not just the \"winner\" — different teams need different frameworks\r\n- Code examples default to Python; mention JS/TS equivalents where available\r\n- For enterprise requirements, flag SOC2 / data residency considerations\r\n- Keep up with the rapidly evolving MCP (Model Context Protocol) and A2A protocol integrations\r\n- Recommend starting simple: single agent → multi-agent only when genuinely needed\n\nFile v1.0.0:_meta.json\n\n{\n  \"ownerId\": \"kn74e704j3ygjcygnpf02rdvd185js13\",\n  \"slug\": \"ai-agent-orchestration-advisor\",\n  \"version\": \"1.0.0\",\n  \"publishedAt\": 1778854315118\n}","readmeExcerpt":"Skill: AI Agent Orchestration Advisor Owner: gechengling Summary: Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework. AI-powered multi-agent framework comparison and selection assistant — analyze use cases, compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture recommendat","codeSnippets":[],"executableExamples":[],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"---\r\nname: AI Agent Orchestration Advisor\r\ndescription: >\r\n  Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework.  AI-powered multi-agent framework comparison and selection assistant — analyze use cases,\r\n  compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture\r\n  recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph,\r\n  CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework\r\n  comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体,\r\n  智能体编排, 框架对比, 智能体架构.\r\nversion: \"3.0.2\"\r\n---\r\n\r\n# AI Agent Orchestration Advisor / 多智能体编排选型顾问\r\n\r\n> Your expert co-pilot for designing, selecting, and implementing multi-agent AI systems.\r\n> 面向多智能体系统的选型、架构设计与落地实施：从需求澄清到框架对比、架构拓扑、起步代码、评测护栏与成本测算。\r\n\r\n## What This Skill Does\r\n\r\nIn 2026, the agentic AI ecosystem exploded — LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Claude Agent SDK, and Strands Agents all compete for developer mindshare. Picking the wrong framework wastes weeks. This skill helps you:\r\n\r\n- **Choose the right framework** for your specific use case (workflow complexity, state management, team size, hosting requirements)\r\n- **Generate architecture diagrams** and data flow specs for multi-agent systems\r\n- **Produce starter code** scaffolds (Python) for the chosen framework\r\n- **Analyze trade-offs** across orchestration patterns (hierarchical, sequential, parallel, event-driven)\r\n- **Design evaluation and guardrails** before go-live, including human-review checkpoints\r\n- **Estimate cost and capacity** per run and per month\r\n- **Debug and optimize** existing multi-agent implementations\r\n\r\n## Trigger Words / 触发词\r\n\r\n**English Triggers:** multi-agent framework comparison, agent orchestration design, LangGraph vs CrewAI, agent architecture topology, agent evaluation and guardrails, agent cost estimation, agent state persistence debugging\r\n\r\n**English Non-Triggers:** single prompt writing, general chatbot building, model fine-tuning, RAG pipeline basics unrelated to agents, web scraping scripts, generic Python debugging\r\n\r\n**中文触发词（须落在多智能体选型或设计任务上才触发）：** 多智能体框架怎么选 / 智能体编排模式有哪些 / 有状态智能体怎么做断点恢复 / 多智能体评测指标怎么定 / 智能体护栏怎么设计 / 多智能体成本怎么测算 / 智能体循环失控怎么处理\r\n\r\n**不触发清单：** 单条提示词优化、通用大模型问答、模型微调与训练、普通脚本编写、纯项目管理咨询\r\n\r\n**路由判定三步**：① 任务是否涉及两个及以上 Agent 的协作、编排或框架选型？② 是否需要输出架构、代码脚手架、评测或成本结论？③ 前两步均为是才启用本技能；单 Agent 场景转提示词工程或应用开发类技能。\r\n\r\n## Target Users / 适用人群\r\n\r\n\r\n- AI engineers building production agent systems\r\n- Data scientists exploring agentic automation\r\n- Product managers scoping agent-based features\r\n- Developers migrating from single LLM to multi-agent pipelines\r\n- **金融行业科技与数据团队**：需同时满足业务交付与合规留痕要求\r\n\r\n---\r\n\r\n## Ecosystem & Compliance Updates [2026-10-09] / 生态与合规动态\r\n\r\n| 类型 | 内容摘要 | 对选型的影响 | 落地动作 | 优先级 |\r\n|-----|---------|-----------|---------|-------|\r\n| 框架成熟度 | LangGraph 状态机路线趋于稳定：状态持久化/长期记忆/错误恢复成为基线能"},{"path":"_meta.json","content":"{\n  \"ownerId\": \"kn74e704j3ygjcygnpf02rdvd185js13\",\n  \"slug\": \"ai-agent-orchestration-advisor\",\n  \"version\": \"3.0.2\",\n  \"publishedAt\": 1791523200414\n}"},{"path":"skill-card.md","content":"## Description:\n\nHelps developers compare multi-agent frameworks and design architectures, starter code, evaluations, and guardrails without installing, running, or deploying frameworks.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[gechengling](https://clawhub.ai/user/gechengling)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nDevelopers and AI engineers use this skill to select multi-agent frameworks and plan agent topology, starter code, evaluation, guardrails, and costs. Product and data teams can use its trade-off analysis to scope production workflows.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Example integrations may send business data to external search or third-party services.\n\nMitigation: Review each proposed tool and use approved internal services where data must remain private.\n\nRisk: Adapting starter code without safeguards may expose secrets or sensitive content in logs.\n\nMitigation: Keep secrets in environment variables, minimize input data, and redact logs before use.\n\nRisk: Framework recommendations and generated code may be unsuitable for a particular production environment.\n\nMitigation: Human-review recommendations and examples, then test them against the team's security, evaluation, and deployment requirements.\n\n## Reference(s):\n\n- [ClawHub skill release](https://clawhub.ai/gechengling/skills/ai-agent-orchestration-advisor)\n\n## Skill Output:\n\n**Output Type(s):** [Text, Markdown, Code, Guidance]\n\n**Output Format:** [Markdown with comparison tables, architecture diagrams, and Python code blocks]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [Advisory output only; example code requires review and adaptation before execution.]\n\n## Skill Version(s):\n\n3.0.2 (source: SKILL.md frontmatter and ClawHub release)\n\n## Ethical Considerations:\n\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment."}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":"Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework. AI-powered multi-agent framework comparison and selection assistant — analyze use cases, compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture recommendations and starter code. Keywords: multi-agent, agent orchestration, LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK, Strands Agents, AutoGen, agent framework comparison, AI agent architecture, multi-agent system design, agent SDK, 多智能体, 智能体编排, 框架对比, 智能体架构. Skill: AI Agent Orchestration Advisor Owner: gechengling Summary: Scope: framework comparison, architecture design, starter-code scaffolds and evaluation design for multi-agent systems; it does not run, install, or deploy any framework. AI-powered multi-agent framework comparison and selection assistant — analyze use cases, compare LangGraph/CrewAI/OpenAI Agents SDK/Claude Agent SDK, generate architecture recommendat","editorialQuality":{"score":100,"threshold":65,"status":"ready","wordCount":1232,"uniquenessScore":47,"reasons":[]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-10T18:27:20.868Z","emptyReason":"No screenshots, media assets, or demo links are available."},"primaryImageUrl":null,"mediaAssetCount":0,"assets":[],"demoUrl":null},"ownerResources":{"evidence":{"source":"unclaimed","verified":false,"confidence":"low","updatedAt":"2026-10-10T18:27:20.868Z","emptyReason":"This page has not been claimed by the agent owner."},"hasCustomPage":false,"customPageUpdatedAt":null,"customLinks":[],"structuredLinks":{"docsUrl":null,"demoUrl":null,"supportUrl":null,"pricingUrl":null,"statusUrl":null},"customPage":null},"relatedAgents":{"evidence":{"source":"protocol-neighbors","verified":false,"confidence":"medium","updatedAt":"2026-10-10T21:49:50.582Z","emptyReason":null},"items":[{"id":"8ebccd8e-3863-4187-8355-c3f14e1f9edf","entityType":"agent","canonicalPath":"/agent/iofficeai-aionui","slug":"iofficeai-aionui","name":"AionUi","description":"Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!","url":"https://github.com/iOfficeAI/AionUi","homepage":"https://www.aionui.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-10-09T19:11:12.944Z","createdAt":"2026-02-25T03:38:16.584Z","downloads":null},{"id":"b917f68a-ebff-438e-84f8-3f4b2494c0bc","entityType":"agent","canonicalPath":"/agent/activepieces-activepieces","slug":"activepieces-activepieces","name":"activepieces","description":"AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents","url":"https://github.com/activepieces/activepieces","homepage":"https://www.activepieces.com","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-15T02:22:12.426Z","createdAt":"2026-02-25T03:38:12.412Z","downloads":null},{"id":"5cb26759-3a39-483f-94cf-276a98c13bb8","entityType":"agent","canonicalPath":"/agent/cherryhq-cherry-studio","slug":"cherryhq-cherry-studio","name":"cherry-studio","description":"AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs","url":"https://github.com/CherryHQ/cherry-studio","homepage":"https://cherry-ai.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-11T14:38:40.986Z","createdAt":"2026-02-25T03:38:19.379Z","downloads":null},{"id":"6f6582d0-5d76-4f0f-b81d-86520247950b","entityType":"agent","canonicalPath":"/agent/copilotkit-copilotkit","slug":"copilotkit-copilotkit","name":"CopilotKit","description":"The Frontend for Agents & Generative UI. React + Angular","url":"https://github.com/CopilotKit/CopilotKit","homepage":"https://docs.copilotkit.ai","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-03-25T09:50:57.846Z","createdAt":"2026-02-25T03:39:14.617Z","downloads":null}],"links":{"hub":"/agent","source":"/agent/source/clawhub","protocols":[{"label":"OpenClaw","href":"/agent/protocol/openclew"}]}}}