[{"data":1,"prerenderedAt":808},["ShallowReactive",2],{"labs":3},[4,155,522,576],{"id":5,"title":6,"author":7,"body":8,"category":138,"date":139,"description":140,"extension":141,"featured":142,"home_position":143,"image":144,"meta":145,"navigation":142,"order":143,"path":146,"seo":147,"status":148,"stem":149,"tags":150,"__hash__":154},"content/labs/ZhiQi.md","ZhiQi","sibuchen",{"type":9,"value":10,"toc":132},"minimark",[11,18,23,28,33,48,52,73,77,121],[12,13,14],"blockquote",{},[15,16,17],"p",{},"PS：ZhiQi (执棋) 如棋盘博弈，先观全局、定谋略 (Plan)，再步步为营、随机应变 (ReAct)。未落子时，全盘局势已在心中推演完毕。它将宏大的困局拆解为一个个精妙的定式，步步为营，运筹帷幄之中，决胜千里之外。♟~",[19,20,22],"h1",{"id":21},"前置知识psplan-and-solve","前置知识：P&S（Plan-and-Solve）",[12,24,25],{},[15,26,27],{},"父Agent + 子Agent（Planner、Executor）",[29,30,32],"h2",{"id":31},"ps过程","P&S过程",[34,35,36,40],"ol",{},[37,38,39],"li",{},"规划阶段 (Planning Phase)： 首先，智能体会接收用户的完整问题。它的第一个任务不是直接去解决问题或调用工具，而是将问题分解，并制定出一个清晰、分步骤的行动计划。这个计划本身就是一次大语言模型的调用产物。",[37,41,42,43],{},"执行阶段 (Solving Phase)： 在获得完整的计划后，智能体进入执行阶段。它会严格按照计划中的步骤，逐一执行。每一步的执行都可能是一次独立的 LLM 调用，或者是对上一步结果的加工处理，直到计划中的所有步骤都完成，最终得出答案。\n",[44,45],"img",{"alt":46,"src":47},"","/assets/ZhiQi/Plan-and-Solve%20%E8%8C%83%E5%BC%8F%E7%9A%84%E4%B8%A4%E9%98%B6%E6%AE%B5%E5%B7%A5%E4%BD%9C%E6%B5%81.png",[29,49,51],{"id":50},"ps的优势","P&S的优势",[34,53,54,61,67],{},[37,55,56,60],{},[57,58,59],"strong",{},"解决盲目性","：通过前置 Planning，避免了 LLM 在处理长任务时因为注意力分散导致的死循环。",[37,62,63,66],{},[57,64,65],{},"高容错率","：每步执行内部依然使用 ReAct，通过 Observation 动态修正局部错误，而不是机械执行计划。",[37,68,69,72],{},[57,70,71],{},"可解释性极强","：终端日志清晰展示了“大计划 -> 小思考 -> 真实行动”的完整链条。",[29,74,76],{"id":75},"ps的劣势","P&S的劣势",[34,78,79,85,91,97,103,109,115],{},[37,80,81,84],{},[57,82,83],{},"对 Planner 要求高","：如果初始计划拆解错误，后续执行可能偏离目标（虽有结果反馈修正，但核心逻辑受限）。",[37,86,87,90],{},[57,88,89],{},"中途无法与用户交互","：Plan制定好后无法更改，无法根据用户的需求生成更合适的计划。",[37,92,93,96],{},[57,94,95],{},"运行效率低","：细分成了许多小步骤，每一步又需要执行完整的ReAct。",[37,98,99,102],{},[57,100,101],{},"上下文记忆爆炸","：大历史+小历史，需要更好的存储方式。例如，ZhiQi 在解决“明天我和父母要从广州出发去邵阳游玩，一共是2天1夜，有什么推荐的景点”问题时，Planner（子Agent）制定的计划一共有10步，Executor（子Agent）每次执行完ReAct 后又有一个 Finish 需要记录。",[37,104,105,108],{},[57,106,107],{},"Token 消耗较高","：因为每个步骤都可能涉及多次 LLM 调用，相比于单次生成，成本与延迟更高。",[37,110,111,114],{},[57,112,113],{},"缺乏审查与反馈机制","：中间某个环节出错/达到最大ReAct限制，Agent 会直接放弃该步骤，导致该步骤在“大历史”中被错误记录 / 被记录为“该步骤已达最大重试次数，未得出结论”，从而影响下一步骤的执行 / 逼迫 LLM 在执行下一步时不得不幻想此步骤的可能结果。例如，ZhiQi 在执行\"步骤 3/10: 搜索邵阳市区及周边核心景点（如崀山、南山牧场、魏源故居等）并筛选适合2天1夜行程的景点组合\"时，由于达到最大循环次数（i=5），迫使 ReAct 终止，导致在步骤 4/10 时出现了“【Thought】: 步骤3（搜索邵阳核心景点）未成功完成，但我需要基于已有信息和常识来推进步骤4。根据步骤1和2的结果，我已知......”",[37,116,117,120],{},[57,118,119],{},"受 LLM 的影响大","：大模型训练数据集的陈旧性。例如，ZhiQi 在执行\"步骤 2/10: 根据交通到达时间确定第一天可游玩的有效时长\"时，制定了错误的搜索参数：“Search【广州南站到邵阳高铁时刻表 2024 早上发车时间】”",[19,122,124,125],{"id":123},"源码地址github","源码地址：",[126,127,131],"a",{"href":128,"rel":129},"https://github.com/sibuchen/ZhiQi--PlanAndSolveAgent",[130],"nofollow","Github",{"title":46,"searchDepth":133,"depth":133,"links":134},2,[135,136,137],{"id":31,"depth":133,"text":32},{"id":50,"depth":133,"text":51},{"id":75,"depth":133,"text":76},"agent","2026-03-28","ZhiQi (执棋) 是一个将 Plan-and-Solve (规划与解决) 范式与 ReAct (推理与行动) 范式深度融合的 AIAgent 架构。它旨在解决传统 ReAct 代理在面对复杂、长期任务时容易“绕路”或“陷入死循环”的问题。","md",true,1,null,{},"/labs/zhiqi",{"title":6,"description":140},"ARCHIVED","labs/ZhiQi",[151,152,153,7],"AIAgent","P&S","ReAct","-Rj8H7wiHTdlKLiOqJQGkKnZCTxXB1UFJncPH-aKyzc",{"id":156,"title":157,"author":7,"body":158,"category":138,"date":513,"description":514,"extension":141,"featured":142,"home_position":515,"image":144,"meta":516,"navigation":142,"order":133,"path":517,"seo":518,"status":148,"stem":519,"tags":520,"__hash__":521},"content/labs/WenJian.md","WenJian",{"type":9,"value":159,"toc":501},[160,165,169,174,178,195,199,202,205,216,220,225,369,373,401,405,439,443,495],[12,161,162],{},[15,163,164],{},"PS：“问”代表推理与探询，“剑”代表行动与决断。WenJian (问剑) 如侠客行走江湖，每遇迷障，先凝神“问”道于心（Reason），随即挥“剑”破局（Action）。剑出必有回响（Observation），回响再引剑招，往复之间，迷雾散尽 🗡~",[19,166,168],{"id":167},"前置知识reactreasoning-and-acting","前置知识：ReAct（Reasoning and Acting）",[12,170,171],{},[15,172,173],{},"推理使得行动更具有目的性，而行动则为推理提供了事实依据。",[29,175,177],{"id":176},"react过程","ReAct过程",[34,179,180,183,186,189],{},[37,181,182],{},"Thought（思考）：内心独白。分析当前情况 分解任务 制定下一步计划 / 反思上一步结果。执行Search函数 传入参数\"sibuchen\"",[37,184,185],{},"Action（行动）：具体动作。调用外部工具 / 输出。Search(\"sibuchen\")",[37,187,188],{},"Observation（观察）：环境变化。从外部工具返回的结果。\"计科学生\"",[37,190,191,192],{},"循环 上下文增加 直到Action=输出\nReAct 范式中的“思考-行动-观察”协同循环图解：\n",[44,193],{"alt":46,"src":194},"/assets/WenJian/ReAct%20%E8%8C%83%E5%BC%8F%E4%B8%AD%E7%9A%84%E2%80%9C%E6%80%9D%E8%80%83-%E8%A1%8C%E5%8A%A8-%E8%A7%82%E5%AF%9F%E2%80%9D%E5%8D%8F%E5%90%8C%E5%BE%AA%E7%8E%AF.png",[29,196,198],{"id":197},"react场景","ReAct场景",[15,200,201],{},"需要调用外部工具API的场景：查询实时信息、搜索专业知识、使用专业工具（计算器、代码解释器）、操作数据库、调用第三方服务API",[29,203,204],{"id":204},"tools的三要素",[34,206,207,210,213],{},[37,208,209],{},"名称：一个简洁、唯一的标识符，供智能体在Action中调用",[37,211,212],{},"描述：一段清晰的自然语言描述，说明这个工具的用途。最为关键。LLM依赖此描述来判断何时调用此工具",[37,214,215],{},"执行逻辑：真正执行的函数/方法",[29,217,219],{"id":218},"wenjian-vs-nova","WenJian VS Nova",[221,222,224],"h3",{"id":223},"️-核心差异-differences","🛠️ 核心差异 (Differences)",[226,227,228,242],"table",{},[229,230,231],"thead",{},[232,233,234,239],"tr",{},[235,236,238],"th",{"align":237},"left","模块",[235,240,241],{"align":237},"WenJian 进阶实现",[243,244,245,263,301],"tbody",{},[232,246,247,256],{},[248,249,250],"td",{"align":237},[57,251,252],{},[253,254,255],"code",{},"config/settings.py",[248,257,258,259,262],{"align":237},"1. 完善的 ",[57,260,261],{},"配置检查"," 验证机制",[232,264,265,272],{},[248,266,267],{"align":237},[57,268,269],{},[253,270,271],{},"llm/client.py",[248,273,274,275,278,279,282,283,286,287,289,290,293,294,296,297,300],{"align":237},"1. ",[57,276,277],{},"自动读取"," setting 配置文件 (无需手动传参) ",[280,281],"br",{}," 2. 严密的 ",[57,284,285],{},"超时控制"," ",[280,288],{}," 3. ",[57,291,292],{},"强制熔断"," 机制 (严防模型幻觉) ",[280,295],{}," 4. ",[57,298,299],{},"流式响应"," 及其异常保护处理",[232,302,303,310],{},[248,304,305],{"align":237},[57,306,307,309],{},[253,308,138],{}," 核心逻辑",[248,311,274,312,315,316,318,319,322,323,289,325,328,329,296,331,334,335,337,338,341,342,344,345,286,348,350,351,354,355,357,358,361,362,364,365,368],{"align":237},[57,313,314],{},"动态工具箱"," (Prompt 实时注入) ",[280,317],{}," 2. ",[57,320,321],{},"单样本 (One-shot)"," 引导逻辑 ",[280,324],{},[57,326,327],{},"工具执行器 (Executor)"," 统一管理 ",[280,330],{},[57,332,333],{},"工厂模式 (Factory)"," 实现配置与逻辑深度解耦 ",[280,336],{}," 5. 支持运行时 ",[57,339,340],{},"动态注册/修改"," 工具 ",[280,343],{}," 6. 规范的 ",[57,346,347],{},"LLM 输出解析器",[280,349],{}," 7. ",[57,352,353],{},"提示词模板拆分"," (System + User 分离响应) ",[280,356],{}," 8. ",[57,359,360],{},"三重熔断"," 安全架构 (Prompt + Client + Core) ",[280,363],{}," 9. 极其清晰的 ",[57,366,367],{},"ReAct 状态机与记忆"," 链路",[221,370,372],{"id":371},"共同基因-similarities","🤝 共同基因 (Similarities)",[374,375,376,383,389,395],"ul",{},[37,377,378,379,382],{},"✅ ",[57,380,381],{},"核心范式",": 均遵循标准 ReAct 推理闭环逻辑",[37,384,378,385,388],{},[57,386,387],{},"容错能力",": 内置标准循环计数 (容错机制 i=5)",[37,390,378,391,394],{},[57,392,393],{},"输出净化",": 自动截断多余的冗余 Thought-Action 对",[37,396,378,397,400],{},[57,398,399],{},"记忆溯源",": 完整支持多轮历史对话上下文记忆",[29,402,404],{"id":403},"react的优势","ReAct的优势",[34,406,407,415,422,432],{},[37,408,409,410],{},"思考（易产生幻觉）+ 行动（易缺乏规划）=  ",[411,412,414],"span",{"style":413},"color:#d83931","相互影响",[37,416,417,418,421],{},"形成”Thought + Action + Observation“链条  -- ",[411,419,420],{"style":413},"Agent所有行为公开透明"," -- 高可解释性 -- 有助于 理解、信任、调试 Agent",[37,423,424,427,428,431],{},[411,425,426],{"style":413},"动态规划与纠错"," -- 没用一次性生成完整计划 而是 ",[411,429,430],{"style":413},"” 走一步，看一步 “"," 每一步的 Observation 都会影响下一步的 Thought 和 Action ，可以实现Agent自我调优。例如，在 WenJian 查询 ” 张雪峰怎么了？“ 时，它认为 Observation 的结果可能是谣言 于是再执行了相同的 Action 并细化了搜索参数。",[37,433,434,435,438],{},"Agent = LLM + Tools + History + Core  --  实现了",[411,436,437],{"style":413},"LLM（亚符号）与外部Tools（符号）的深度结合"," 有效避免了LLM幻觉（例如 计算 解析 搜索任务）",[29,440,442],{"id":441},"react的劣势","ReAct的劣势",[34,444,445,459,469,486],{},[37,446,447,450,451,454,455],{},[411,448,449],{"style":413},"过度依赖LLM ，LLM的逻辑推理能力直接影响 Thought 的有效/正确规划，LLM的指令遵循能力与格式化输出能力直接影响 Action 的有效性。","这就是为什么 WenJian 要在 core 中实现对各自可能错误的”驳回“。甚至",[411,452,453],{"style":413}," Agent 的效果会受到 LLM 训练时的数据集影响","，例如 WenJian 在搜索 ”张雪峰怎么了？“ 时，LLM 出现了 ”2026是未来时间，该消息是谣言“ 的误判。",[411,456,458],{"style":457},"color:#4f81bd","--  尝试不同的模型/参数",[37,460,461,464,465,468],{},[411,462,463],{"style":413},"执行效率低下","，每次 Thought 只规划一步（Action），每次得到 Observation 才进行下一次 Thought 。一个任务需要",[411,466,467],{"style":413},"多次执行串行的ReAct循环，需要多次调用 LLM ","，需要消耗大量的时间。",[37,470,471,474,475,478,479,482,483],{},[411,472,473],{"style":413},"提示词的脆弱性，整个机制建立在一个精选设计的提示词模板上，模板中任何一个微小的变化，设置是用词的差异，都会影响 LLM 的行为。","例如，在prompt中去掉对Observation的描述（\"你当前正处于一个 ",[57,476,477],{},"Thought -> Action -> Observation"," 的闭环决策链中\"） 可以大幅度降低幻觉。另外，",[411,480,481],{"style":413},"由于 Thought 中心化堆积在一个 LLM 中，导致在处理复杂的任务时提示词容易失效，LLM 出现越权行为。","例如 WenJian 在处理 ” 邵阳有哪些好玩的地方？“时本意是想先搜索邵阳的热门景点，再搜索各个景点的具体介绍与推荐，但是由于上下文的堆砌，导致 LLM 直接幻想出了 Observation 的内容，并自我迭代”Thought + Action + Observation“链条。",[411,484,485],{"style":457},"--  为promot添加少样本",[37,487,488,491,492],{},[411,489,490],{"style":413},"不适合需要规划的长期任务，步进式决策模式使得 Agent 缺乏一个全局的、长远的规划，容易造成绕远路/原地打转的循环当中。","例如 WenJian 在处理 ”张雪峰怎么了？“ 时，出现时间的误判后就执着于搜索 ” 张雪峰社交媒体账号头像的颜色 “ 在无关的搜索中陷入了死循环。",[411,493,494],{"style":457},"--  打印完整 ReAct 流程进行分析",[19,496,124,497],{"id":123},[126,498,131],{"href":499,"rel":500},"https://github.com/sibuchen/WenJian--ReActAgent",[130],{"title":46,"searchDepth":133,"depth":133,"links":502},[503,504,505,506,511,512],{"id":176,"depth":133,"text":177},{"id":197,"depth":133,"text":198},{"id":204,"depth":133,"text":204},{"id":218,"depth":133,"text":219,"children":507},[508,510],{"id":223,"depth":509,"text":224},3,{"id":371,"depth":509,"text":372},{"id":403,"depth":133,"text":404},{"id":441,"depth":133,"text":442},"2026-03-25","WenJian (问剑) 是一个高度解耦、易于扩展的轻量级 AIAgent 框架。它基于经典的 ReAct (Reason + Action) 范式实现，允许大语言模型（LLM）在解决复杂问题时，通过“思考（Thought）- 行动（Action）- 观察（Observation）”的循环，自主调用外部工具获取信息并得出最终结论。",4,{},"/labs/wenjian",{"title":157,"description":514},"labs/WenJian",[151,153,7],"ALLdwd0GY86jhyyPXbtjOtxgNokfK0gNnS3dIKjgmdo",{"id":523,"title":524,"author":7,"body":525,"category":138,"date":568,"description":569,"extension":141,"featured":142,"home_position":133,"image":144,"meta":570,"navigation":142,"order":509,"path":571,"seo":572,"status":148,"stem":573,"tags":574,"__hash__":575},"content/labs/Nova.md","Nova",{"type":9,"value":526,"toc":566},[527,532,536,547,558],[12,528,529],{},[15,530,531],{},"Nova （诺瓦）在拉丁语意为“新的”，天文学中指“新星”。每一次旅行都是新的开始，发现新的风景~",[19,533,535],{"id":534},"前置知识智能体的构成与运行原理","前置知识：智能体的构成与运行原理",[15,537,538,539,543,544],{},"智能体并非一次性完成任务，而是通过一个持续的循环与环境进行交互，这个核心机制被称为",[411,540,542],{"style":541},"color:red","智能体循环 (Agent Loop)","：\n",[44,545],{"alt":46,"src":546},"/assets/Nova/%E6%99%BA%E8%83%BD%E4%BD%93%E4%B8%8E%E7%8E%AF%E5%A2%83%E4%BA%A4%E4%BA%92%E7%9A%84%E5%9F%BA%E6%9C%AC%E5%BE%AA%E7%8E%AF.png",[15,548,549,550,553,554,557],{},"在工程实践中，为了让 LLM 能够有效驱动这个循环，我们需要一套明确的",[411,551,552],{"style":541},"交互协议 (Interaction Protocol)","来规范其与环境之间的信息交换。\n在许多现代智能体框架中，这一协议体现在对智能体每一次输出的结构化定义上。智能体的输出不再是单一的自然语言回复，而是一段遵循特定格式的文本，其中明确地展示了其内部的推理过程与最终决策。\n由于该循环结构通常由 Thought、Action、Observation 三个部分构成，因此被称为",[411,555,556],{"style":541},"Thought-Action-Observation交互范式","。通过它，LLM智能体得以将内部的语言推理能力，与外部环境的真实信息和工具操作能力有效地结合起来。",[19,559,561,562],{"id":560},"源码位置github","源码位置：",[126,563,131],{"href":564,"rel":565},"https://github.com/sibuchen/Nova--TravelAgent",[130],{"title":46,"searchDepth":133,"depth":133,"links":567},[],"2026-03-20","Nova 是一个由作者设计的简单Agent Demo，集中演示了基于 Thought-Action-Observation 循环的智能体所具备的四项基本能力：任务分解、工具调用、上下文理解和结果合成。通过这种循环的不断迭代，Nova 得以将一个模糊的用户意图，转化为一系列具体、可执行的步骤，并最终达成目标。",{},"/labs/nova",{"title":524,"description":569},"labs/Nova",[151,153,7],"rMZtyutqD2YFPZZkfSXZuFxCwNcPIDVEMCC6ipKjQS8",{"id":577,"title":578,"author":7,"body":579,"category":138,"date":795,"description":796,"extension":141,"featured":797,"home_position":144,"image":144,"meta":798,"navigation":142,"order":515,"path":799,"seo":800,"status":801,"stem":802,"tags":803,"__hash__":807},"content/labs/CoPaw.md","CoPaw",{"type":9,"value":580,"toc":787},[581,584,587,601,604,611,634,645,648,652,661,672,682,685,692,697,700,704,709,717,722,726,731,739,744,748,752,757,761,772],[19,582,583],{"id":583},"安全知识",[15,585,586],{},"CoPaw = OpenClaw + 阿里云（阿里云通义实验室 -- AgentScope 团队）",[34,588,589,595,598],{},[37,590,591,592],{},"沙盒 -- 备用机/Docker/VMware(快照功能)",[44,593],{"alt":46,"src":594},"/assets/CoPaw/VMware(%E5%BF%AB%E7%85%A7%E5%8A%9F%E8%83%BD).png",[37,596,597],{},"隔离账号/密钥 （API）",[37,599,600],{},"Prompt MCPs Skills 审查",[19,602,603],{"id":603},"安装部署",[15,605,606],{},[126,607,610],{"href":608,"rel":609},"https://copaw.agentscope.io/",[130],"CoPaw官网",[34,612,613,616,619],{},[37,614,615],{},"命令1-需要科学上网/网络稳定的时间地点 （有些安装资源需要连接Github）",[37,617,618],{},"命令2-重开一个终端（刷新终端快照）",[37,620,621,622,625,626,629,630,633],{},"命令3-一定不要开网络代理 （目前版本的CoPaw客户端还没有处理SOCKS代理的相关配置，依赖的底层 HTTP 库（",[253,623,624],{},"httpx","）",[57,627,628],{},"不支持"," ",[253,631,632],{},"socks://"," 这种协议前缀。）",[12,635,636],{},[15,637,638,639,644],{},"特殊问题",[126,640,643],{"href":641,"rel":642},"https://copaw.agentscope.io/docs/quickstart",[130],"查看官方文档"," / 询问LLM（推荐Qwen）",[19,646,647],{"id":647},"基础使用",[29,649,651],{"id":650},"连接llm","连接LLM",[15,653,654,655,660],{},"网址推荐：",[126,656,659],{"href":657,"rel":658},"https://www.aliyun.com/product/bailian",[130],"阿里云百炼平台","\nLLM-API设置：",[34,662,663,666,669],{},[37,664,665],{},"模型提供商地址 - 请求行",[37,667,668],{},"API Key - 请求头",[37,670,671],{},"模型名 - 请求头",[12,673,674],{},[15,675,676,677],{},"参考视频：",[126,678,681],{"href":679,"rel":680},"https://www.douyin.com/video/7615978059157622067",[130],"科技小辛",[29,683,684],{"id":684},"连接第三方频道",[15,686,687,691],{},[126,688,643],{"href":689,"rel":690},"https://copaw.agentscope.io/docs/channels",[130]," -- 很详细 无脑跟",[12,693,694],{},[15,695,696],{},"推荐钉钉 / 飞书",[19,698,699],{"id":699},"高级使用",[29,701,703],{"id":702},"mcpserver","MCPserver",[12,705,706],{},[15,707,708],{},"MCPs = MCP + API = 获取数据 = 能干活",[15,710,711,712],{},"推荐网址：",[126,713,716],{"href":714,"rel":715},"https://smithery.ai/",[130],"smithery.ai",[12,718,719],{},[15,720,721],{},"注意：选择官方的 星标多的 热门的",[29,723,725],{"id":724},"skills","Skills",[12,727,728],{},[15,729,730],{},"Skills = .md + .py = 渐进式提示词 = 干好活",[15,732,733,734],{},"推荐检测网址：",[126,735,738],{"href":736,"rel":737},"https://ai.gendigital.com/skill-scanner",[130],"Gen",[12,740,741],{},[15,742,743],{},"Skills的网址可以使用OpenClaw的",[19,745,747],{"id":746},"agents","Agents",[29,749,751],{"id":750},"agent的分类","Agent的分类",[15,753,754],{},[44,755],{"alt":46,"src":756},"/assets/CoPaw/Agent%E7%9A%84%E5%88%86%E7%B1%BB.png",[29,758,760],{"id":759},"多agent","多Agent",[15,762,763,767,768,771],{},[126,764,643],{"href":765,"rel":766},"https://copaw.agentscope.io/docs/multi-agent",[130]," -- 很详细\n版本要求：",[57,769,770],{},"v0.1.0","\n常见问题：",[34,773,774,777,784],{},[37,775,776],{},"如何更新 -- 重新执行一遍安装的三个命令",[37,778,779,780,783],{},"打开网页标识为v0.1.0 但是未看到多Agent按钮 -- 清空浏览器缓存（同时按下 ",[253,781,782],{},"Ctrl + Shift + Delete"," 键）",[37,785,786],{},"能否实现多Agent协作 -- 目前版本不能 官方发布会说预期在下一个版本发布",{"title":46,"searchDepth":133,"depth":133,"links":788},[789,790,791,792,793,794],{"id":650,"depth":133,"text":651},{"id":684,"depth":133,"text":684},{"id":702,"depth":133,"text":703},{"id":724,"depth":133,"text":725},{"id":750,"depth":133,"text":751},{"id":759,"depth":133,"text":760},"2026-03-16","阿里云推出的一款个人智能助理（AI Agent）框架，旨在为用户提供可本地部署或云端部署的专属 AI 数字员工。它支持接入钉钉、飞书等通讯工具，具备记忆系统（ReMe）和可插拔技能（Skills）架构。",false,{},"/labs/copaw",{"title":578,"description":796},"ACTIVE","labs/CoPaw",[151,804,805,806],"开源框架","阿里云","OpenClaw","nxLK1QTFPS5H90h0WT9WJZloeOAnR1qKUOyMVxmeDBk",1774960320850]