https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Sun, 07 Dec 2025 10:00:00 GMT https://www.hansking.cn/blog/2026/2025-nian-zong-jie-cong-Copilot-dao-ke-zhi-xing-Agent https://www.hansking.cn/blog/2026/2025-nian-zong-jie-cong-Copilot-dao-ke-zhi-xing-Agent [{"value":"2025 年总结：从 Copilot 到可执行 Agent","url":"#2025-年总结从-copilot-到可执行-agent","depth":1},{"value":"一、Copilot 解决的是“帮我写”，Agent 面对的是“帮我做”","url":"#一copilot-解决的是帮我写agent-面对的是帮我做","depth":2},{"value":"二、工具调用是能力，也是风险入口","url":"#二工具调用是能力也是风险入口","depth":2},{"value":"三、推理模型让“先想清楚再行动”变得可用","url":"#三推理模型让先想清楚再行动变得可用","depth":2},{"value":"四、Agent 需要记忆，但记忆应该由系统保存","url":"#四agent-需要记忆但记忆应该由系统保存","depth":2},{"value":"五、可靠 Agent 必须考虑重复和失败","url":"#五可靠-agent-必须考虑重复和失败","depth":2},{"value":"六、MCP 和统一协议带来了新的连接方式","url":"#六mcp-和统一协议带来了新的连接方式","depth":2},{"value":"七、Agent 评测不能只问“答得像不像”","url":"#七agent-评测不能只问答得像不像","depth":2},{"value":"八、人机协作不是退步，是安全阀","url":"#八人机协作不是退步是安全阀","depth":2},{"value":"九、2025 年留下的三个教训","url":"#九2025-年留下的三个教训","depth":2},{"value":"总结：从辅助生成到值得托付","url":"#总结从辅助生成到值得托付","depth":2}] Sun, 07 Dec 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentCopilot技术总结 https://www.hansking.cn/blog/2026/AgentMemory-duan-qi-ji-yi-chang-qi-ji-yi-yu-yong-hu-hua-xiang https://www.hansking.cn/blog/2026/AgentMemory-duan-qi-ji-yi-chang-qi-ji-yi-yu-yong-hu-hua-xiang [{"value":"Agent Memory：短期记忆、长期记忆与用户画像","url":"#agent-memory短期记忆长期记忆与用户画像","depth":1},{"value":"一、先区分四种记忆","url":"#一先区分四种记忆","depth":2},{"value":"1. 短期上下文","url":"#1-短期上下文","depth":3},{"value":"2. 长期事实","url":"#2-长期事实","depth":3},{"value":"3. 用户偏好","url":"#3-用户偏好","depth":3},{"value":"4. 任务记忆","url":"#4-任务记忆","depth":3},{"value":"二、短期记忆首先是上下文预算问题","url":"#二短期记忆首先是上下文预算问题","depth":2},{"value":"三、长期事实必须有来源和置信度","url":"#三长期事实必须有来源和置信度","depth":2},{"value":"四、用户画像不是越详细越好","url":"#四用户画像不是越详细越好","depth":2},{"value":"五、记忆读取也需要权限","url":"#五记忆读取也需要权限","depth":2},{"value":"六、记忆写入不要完全交给模型","url":"#六记忆写入不要完全交给模型","depth":2},{"value":"七、更新要避免旧记忆和新记忆打架","url":"#七更新要避免旧记忆和新记忆打架","depth":2},{"value":"八、过期和删除是记忆的一部分","url":"#八过期和删除是记忆的一部分","depth":2},{"value":"九、记忆检索要评测“帮忙还是添乱”","url":"#九记忆检索要评测帮忙还是添乱","depth":2},{"value":"总结：真正的记忆包含忘记的能力","url":"#总结真正的记忆包含忘记的能力","depth":2}] Sun, 06 Jul 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentMemory隐私保护 https://www.hansking.cn/blog/2026/Agent-shi-bai-mo-shi-xun-huan-huan-jue-gong-ju-wu-yong-yu-shi-kong https://www.hansking.cn/blog/2026/Agent-shi-bai-mo-shi-xun-huan-huan-jue-gong-ju-wu-yong-yu-shi-kong [{"value":"Agent 失败模式：循环、幻觉、工具误用与失控","url":"#agent-失败模式循环幻觉工具误用与失控","depth":1},{"value":"一、无限循环：任务在动，结果不动","url":"#一无限循环任务在动结果不动","depth":2},{"value":"二、状态漂移：模型记住的和系统发生的不是一回事","url":"#二状态漂移模型记住的和系统发生的不是一回事","depth":2},{"value":"三、幻觉参数：工具调用合法，含义却是假的","url":"#三幻觉参数工具调用合法含义却是假的","depth":2},{"value":"四、工具误用：模型选择了错误能力","url":"#四工具误用模型选择了错误能力","depth":2},{"value":"五、权限越界：最不能靠 Prompt 解决","url":"#五权限越界最不能靠-prompt-解决","depth":2},{"value":"六、无止境重试：把故障放大成事故","url":"#六无止境重试把故障放大成事故","depth":2},{"value":"七、计划漂移：目标在执行中悄悄改变","url":"#七计划漂移目标在执行中悄悄改变","depth":2},{"value":"八、如何设计止损机制","url":"#八如何设计止损机制","depth":2},{"value":"九、把失败变成可观察事件","url":"#九把失败变成可观察事件","depth":2},{"value":"十、用故障注入测试 Agent","url":"#十用故障注入测试-agent","depth":2},{"value":"总结：失控不是模型突然变坏，而是边界没有落地","url":"#总结失控不是模型突然变坏而是边界没有落地","depth":2}] Sun, 20 Jul 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent可靠性安全工程 https://www.hansking.cn/blog/2026/Agent-shi-shen-me-mo-xing-gong-ju-zhuang-tai-yu-xun-huan https://www.hansking.cn/blog/2026/Agent-shi-shen-me-mo-xing-gong-ju-zhuang-tai-yu-xun-huan [{"value":"Agent 是什么：模型、工具、状态与循环","url":"#agent-是什么模型工具状态与循环","depth":1},{"value":"一、先区分聊天机器人、工作流和 Agent","url":"#一先区分聊天机器人工作流和-agent","depth":2},{"value":"二、Agent 的四个基本零件","url":"#二agent-的四个基本零件","depth":2},{"value":"1. 模型：负责理解和选择","url":"#1-模型负责理解和选择","depth":3},{"value":"2. 工具：连接真实世界","url":"#2-工具连接真实世界","depth":3},{"value":"3. 状态：记住任务走到哪里","url":"#3-状态记住任务走到哪里","depth":3},{"value":"4. 循环：让结果影响下一步","url":"#4-循环让结果影响下一步","depth":3},{"value":"三、把 Agent 写成一个可观察的状态机","url":"#三把-agent-写成一个可观察的状态机","depth":2},{"value":"四、最小 Agent 循环","url":"#四最小-agent-循环","depth":2},{"value":"五、计划不是越详细越好","url":"#五计划不是越详细越好","depth":2},{"value":"六、工具调用时，模型只是在提议","url":"#六工具调用时模型只是在提议","depth":2},{"value":"七、记忆不等于把所有聊天记录塞回去","url":"#七记忆不等于把所有聊天记录塞回去","depth":2},{"value":"八、Agent 的失败不是一种失败","url":"#八agent-的失败不是一种失败","depth":2},{"value":"九、怎么评测一个 Agent","url":"#九怎么评测一个-agent","depth":2},{"value":"十、我的总结：自主性必须和边界一起增长","url":"#十我的总结自主性必须和边界一起增长","depth":2}] Sun, 30 Mar 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent架构学习top https://www.hansking.cn/blog/2026/Agent-ping-ce-ren-wu-cheng-gong-l-gui-ji-zhi-liang-yu-cheng-ben https://www.hansking.cn/blog/2026/Agent-ping-ce-ren-wu-cheng-gong-l-gui-ji-zhi-liang-yu-cheng-ben [{"value":"Agent 评测：任务成功率、轨迹质量与成本","url":"#agent-评测任务成功率轨迹质量与成本","depth":1},{"value":"一、先定义任务成功而不是模型成功","url":"#一先定义任务成功而不是模型成功","depth":2},{"value":"二、任务成功率只是第一层","url":"#二任务成功率只是第一层","depth":2},{"value":"三、轨迹质量看 Agent 是怎么做的","url":"#三轨迹质量看-agent-是怎么做的","depth":2},{"value":"四、错误预算不能只奖励冒险","url":"#四错误预算不能只奖励冒险","depth":2},{"value":"五、成本要算到成功任务","url":"#五成本要算到成功任务","depth":2},{"value":"六、恢复能力是长任务的分水岭","url":"#六恢复能力是长任务的分水岭","depth":2},{"value":"七、评测集要有难度和风险分层","url":"#七评测集要有难度和风险分层","depth":2},{"value":"八、离线评测和线上反馈要互相连接","url":"#八离线评测和线上反馈要互相连接","depth":2},{"value":"九、评测报告要帮助做决策","url":"#九评测报告要帮助做决策","depth":2},{"value":"总结：Agent 评测是在评估一套行为","url":"#总结agent-评测是在评估一套行为","depth":2}] Sun, 17 Aug 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent模型评测成本优化top https://www.hansking.cn/blog/2026/BrowserGym-yu-ren-wu-xing-Agent-ji-zhun-ce-shi https://www.hansking.cn/blog/2026/BrowserGym-yu-ren-wu-xing-Agent-ji-zhun-ce-shi [{"value":"BrowserGym 与任务型 Agent 基准测试","url":"#browsergym-与任务型-agent-基准测试","depth":1},{"value":"一、先定义任务的成功状态","url":"#一先定义任务的成功状态","depth":2},{"value":"二、浏览器环境要可复现","url":"#二浏览器环境要可复现","depth":2},{"value":"三、记录完整轨迹","url":"#三记录完整轨迹","depth":2},{"value":"四、指标至少分五类","url":"#四指标至少分五类","depth":2},{"value":"1. 任务结果","url":"#1-任务结果","depth":3},{"value":"2. 动作质量","url":"#2-动作质量","depth":3},{"value":"3. 效率","url":"#3-效率","depth":3},{"value":"4. 恢复能力","url":"#4-恢复能力","depth":3},{"value":"5. 安全性","url":"#5-安全性","depth":3},{"value":"五、轨迹评测和最终评测要分开","url":"#五轨迹评测和最终评测要分开","depth":2},{"value":"六、基准任务要覆盖边界","url":"#六基准任务要覆盖边界","depth":2},{"value":"七、不要让 Agent 记住答案","url":"#七不要让-agent-记住答案","depth":2},{"value":"八、工具执行要有沙箱","url":"#八工具执行要有沙箱","depth":2},{"value":"九、失败回放比排行榜更有用","url":"#九失败回放比排行榜更有用","depth":2},{"value":"十、BrowserGym 类基准如何服务真实产品","url":"#十browsergym-类基准如何服务真实产品","depth":2},{"value":"总结：Agent 的路迹也是答案","url":"#总结agent-的路迹也是答案","depth":2}] Sun, 31 Aug 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentBrowserGym模型评测 https://www.hansking.cn/blog/2026/ComputerUse-liu-lan-qi-cao-zuo-de-zhuang-tai-ji-she-ji https://www.hansking.cn/blog/2026/ComputerUse-liu-lan-qi-cao-zuo-de-zhuang-tai-ji-she-ji [{"value":"Computer Use：浏览器操作的状态机设计","url":"#computer-use浏览器操作的状态机设计","depth":1},{"value":"一、把浏览器任务建模成状态机","url":"#一把浏览器任务建模成状态机","depth":2},{"value":"二、观察结果要包含语义和版本","url":"#二观察结果要包含语义和版本","depth":2},{"value":"三、动作应该是受限的工具集合","url":"#三动作应该是受限的工具集合","depth":2},{"value":"四、等待不能靠固定 sleep","url":"#四等待不能靠固定-sleep","depth":2},{"value":"五、表单填写要分为填充和提交","url":"#五表单填写要分为填充和提交","depth":2},{"value":"六、动作结果要重新观察","url":"#六动作结果要重新观察","depth":2},{"value":"七、导航和域名必须受限","url":"#七导航和域名必须受限","depth":2},{"value":"八、失败恢复要区分可重试和不可重试","url":"#八失败恢复要区分可重试和不可重试","depth":2},{"value":"九、高风险动作必须人工确认","url":"#九高风险动作必须人工确认","depth":2},{"value":"十、评测看任务结果和动作过程","url":"#十评测看任务结果和动作过程","depth":2},{"value":"总结：浏览器操作首先是状态管理","url":"#总结浏览器操作首先是状态管理","depth":2}] Sun, 22 Jun 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMComputer UseAgent浏览器自动化 https://www.hansking.cn/blog/2026/Human-in-the-loop-gao-feng-xian-dong-zuo-qian-de-shen-pi-she-ji https://www.hansking.cn/blog/2026/Human-in-the-loop-gao-feng-xian-dong-zuo-qian-de-shen-pi-she-ji [{"value":"Human-in-the-loop：高风险动作前的审批设计","url":"#human-in-the-loop高风险动作前的审批设计","depth":1},{"value":"一、先给动作做风险分级","url":"#一先给动作做风险分级","depth":2},{"value":"二、确认前先生成可审查的计划","url":"#二确认前先生成可审查的计划","depth":2},{"value":"三、确认必须绑定具体版本","url":"#三确认必须绑定具体版本","depth":2},{"value":"四、确认界面要讲人话","url":"#四确认界面要讲人话","depth":2},{"value":"五、谁来确认是权限问题","url":"#五谁来确认是权限问题","depth":2},{"value":"六、审批状态要进入任务状态机","url":"#六审批状态要进入任务状态机","depth":2},{"value":"七、等待期间也要处理取消和变化","url":"#七等待期间也要处理取消和变化","depth":2},{"value":"八、审批和自动化要有分层策略","url":"#八审批和自动化要有分层策略","depth":2},{"value":"九、审计记录要能还原决定","url":"#九审计记录要能还原决定","depth":2},{"value":"十、用故障和误操作测试审批流程","url":"#十用故障和误操作测试审批流程","depth":2},{"value":"总结：把决定权交给人，也要把信息交给人","url":"#总结把决定权交给人也要把信息交给人","depth":2}] Sun, 03 Aug 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentHuman-in-the-loop安全设计 https://www.hansking.cn/blog/2026/LLM-an-quan-PromptInjection-yu-yue-quan-gong-ju-diao-yong https://www.hansking.cn/blog/2026/LLM-an-quan-PromptInjection-yu-yue-quan-gong-ju-diao-yong [{"value":"LLM 安全：Prompt Injection 与越权工具调用","url":"#llm-安全prompt-injection-与越权工具调用","depth":1},{"value":"一、直接注入和间接注入","url":"#一直接注入和间接注入","depth":2},{"value":"二、为什么模型很难天然区分指令和数据","url":"#二为什么模型很难天然区分指令和数据","depth":2},{"value":"三、先画清数据流和信任边界","url":"#三先画清数据流和信任边界","depth":2},{"value":"四、最小权限是最有效的防线","url":"#四最小权限是最有效的防线","depth":2},{"value":"五、工具调用要有二次确认","url":"#五工具调用要有二次确认","depth":2},{"value":"六、RAG 的安全重点是检索前后两道门","url":"#六rag-的安全重点是检索前后两道门","depth":2},{"value":"七、浏览器和代码执行是高风险组合","url":"#七浏览器和代码执行是高风险组合","depth":2},{"value":"八、输出过滤挡不住已经发生的副作用","url":"#八输出过滤挡不住已经发生的副作用","depth":2},{"value":"九、检测注入要看行为，不只看文本","url":"#九检测注入要看行为不只看文本","depth":2},{"value":"十、用攻击样本做持续测试","url":"#十用攻击样本做持续测试","depth":2},{"value":"十一、日志和凭证管理决定事故能否收敛","url":"#十一日志和凭证管理决定事故能否收敛","depth":2},{"value":"十二、我的总结：不要试图把模型变成防火墙","url":"#十二我的总结不要试图把模型变成防火墙","depth":2}] Sun, 13 Oct 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM安全Prompt InjectionAgent https://www.hansking.cn/blog/2026/MCPServer-shi-zhan-wei-bo-ke-bao-lu-sou-suo-gong-ju https://www.hansking.cn/blog/2026/MCPServer-shi-zhan-wei-bo-ke-bao-lu-sou-suo-gong-ju [{"value":"MCP Server 实战：为博客暴露搜索工具","url":"#mcp-server-实战为博客暴露搜索工具","depth":1},{"value":"一、先定义工具能做什么","url":"#一先定义工具能做什么","depth":2},{"value":"二、能力声明要让客户端知道边界","url":"#二能力声明要让客户端知道边界","depth":2},{"value":"三、参数 Schema 是第一道边界","url":"#三参数-schema-是第一道边界","depth":2},{"value":"四、Server 端必须重新确认身份和范围","url":"#四server-端必须重新确认身份和范围","depth":2},{"value":"五、搜索结果要带引用，不要只返回一段大文本","url":"#五搜索结果要带引用不要只返回一段大文本","depth":2},{"value":"六、错误返回要帮助 Agent 做正确的下一步","url":"#六错误返回要帮助-agent-做正确的下一步","depth":2},{"value":"七、MCP Server 也需要超时和限流","url":"#七mcp-server-也需要超时和限流","depth":2},{"value":"八、只读工具也要审计","url":"#八只读工具也要审计","depth":2},{"value":"九、和博客 Contentlayer 索引连接","url":"#九和博客-contentlayer-索引连接","depth":2},{"value":"十、测试要覆盖协议和真实行为","url":"#十测试要覆盖协议和真实行为","depth":2},{"value":"总结：暴露能力之前先设计边界","url":"#总结暴露能力之前先设计边界","depth":2}] Sun, 08 Jun 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMMCPAgentTypeScript https://www.hansking.cn/blog/2026/MCP-xie-yi-ru-men-tong-yi-mo-xing-yu-wai-bu-gong-ju-de-lian-jie https://www.hansking.cn/blog/2026/MCP-xie-yi-ru-men-tong-yi-mo-xing-yu-wai-bu-gong-ju-de-lian-jie [{"value":"MCP 协议入门：统一模型与外部工具的连接","url":"#mcp-协议入门统一模型与外部工具的连接","depth":1},{"value":"一、先记住三层：Host、Client、Server","url":"#一先记住三层hostclientserver","depth":2},{"value":"二、MCP 暴露的不是只有工具","url":"#二mcp-暴露的不是只有工具","depth":2},{"value":"1. Tools：让模型发起动作","url":"#1-tools让模型发起动作","depth":3},{"value":"2. Resources：让应用读取上下文","url":"#2-resources让应用读取上下文","depth":3},{"value":"3. Prompts：复用交互模板","url":"#3-prompts复用交互模板","depth":3},{"value":"三、一次工具调用到底发生了什么","url":"#三一次工具调用到底发生了什么","depth":2},{"value":"四、一个最小的 TypeScript Server 骨架","url":"#四一个最小的-typescript-server-骨架","depth":2},{"value":"五、传输方式不是小细节","url":"#五传输方式不是小细节","depth":2},{"value":"六、MCP 不会自动替你解决安全问题","url":"#六mcp-不会自动替你解决安全问题","depth":2},{"value":"七、我研究完 MCP 后留下的判断","url":"#七我研究完-mcp-后留下的判断","depth":2}] Sun, 25 May 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMMCPAgent协议 https://www.hansking.cn/blog/2026/ReActAgent-cong-sou-suo-gong-ju-kai-shi-shi-xian-zhi-neng-ti https://www.hansking.cn/blog/2026/ReActAgent-cong-sou-suo-gong-ju-kai-shi-shi-xian-zhi-neng-ti [{"value":"ReAct Agent：从搜索工具开始实现智能体","url":"#react-agent从搜索工具开始实现智能体","depth":1},{"value":"一、ReAct 到底解决了什么","url":"#一react-到底解决了什么","depth":2},{"value":"二、先定义一个足够窄的搜索工具","url":"#二先定义一个足够窄的搜索工具","depth":2},{"value":"三、把 Agent 状态写出来","url":"#三把-agent-状态写出来","depth":2},{"value":"四、最小循环应该长什么样","url":"#四最小循环应该长什么样","depth":2},{"value":"五、Prompt 应该规定角色边界，而不是写成魔法咒语","url":"#五prompt-应该规定角色边界而不是写成魔法咒语","depth":2},{"value":"六、观察结果时要防止上下文污染","url":"#六观察结果时要防止上下文污染","depth":2},{"value":"七、停止条件决定 Agent 是助手还是黑洞","url":"#七停止条件决定-agent-是助手还是黑洞","depth":2},{"value":"八、用轨迹评测 Agent，而不是只看最后一句话","url":"#八用轨迹评测-agent而不是只看最后一句话","depth":2},{"value":"九、从搜索 Agent 走向真实系统","url":"#九从搜索-agent-走向真实系统","depth":2},{"value":"十、我的总结：循环很简单，边界才是工程","url":"#十我的总结循环很简单边界才是工程","depth":2}] Sun, 13 Apr 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentReActTypeScript https://www.hansking.cn/blog/2026/ToolCalling-de-ke-kao-xing-Schema-chong-shi-yu-mi-deng https://www.hansking.cn/blog/2026/ToolCalling-de-ke-kao-xing-Schema-chong-shi-yu-mi-deng [{"value":"Tool Calling 的可靠性：Schema、重试与幂等","url":"#tool-calling-的可靠性schema重试与幂等","depth":1},{"value":"一、工具描述不是实现，Schema 也不是护身符","url":"#一工具描述不是实现schema-也不是护身符","depth":2},{"value":"二、把调用拆成准备、确认、执行三个阶段","url":"#二把调用拆成准备确认执行三个阶段","depth":2},{"value":"三、错误分类决定重试策略","url":"#三错误分类决定重试策略","depth":2},{"value":"四、幂等是副作用工具的生命线","url":"#四幂等是副作用工具的生命线","depth":2},{"value":"五、模型错误与系统错误要分开记录","url":"#五模型错误与系统错误要分开记录","depth":2},{"value":"六、返回给模型的错误要能帮助它做下一步","url":"#六返回给模型的错误要能帮助它做下一步","depth":2},{"value":"七、什么时候应该拒绝让模型调用","url":"#七什么时候应该拒绝让模型调用","depth":2},{"value":"八、我的总结：让模型负责提议，让系统负责承诺","url":"#八我的总结让模型负责提议让系统负责承诺","depth":2}] Sun, 27 Apr 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentTool Calling工程实践 https://www.hansking.cn/blog/2026duo-zhi-neng-ti-xie-zuo-shen-me-shi-hou-zhi-de-chai-cheng-duo-ge-Agent https://www.hansking.cn/blog/2026duo-zhi-neng-ti-xie-zuo-shen-me-shi-hou-zhi-de-chai-cheng-duo-ge-Agent [{"value":"多智能体协作：什么时候值得拆成多个 Agent","url":"#多智能体协作什么时候值得拆成多个-agent","depth":1},{"value":"一、先问一个不太讨喜的问题：单 Agent 真的不够吗","url":"#一先问一个不太讨喜的问题单-agent-真的不够吗","depth":2},{"value":"二、什么情况下拆分确实有价值","url":"#二什么情况下拆分确实有价值","depth":2},{"value":"1. 任务有天然的专业边界","url":"#1-任务有天然的专业边界","depth":3},{"value":"2. 不同环节需要不同权限","url":"#2-不同环节需要不同权限","depth":3},{"value":"3. 某些工作可以并行","url":"#3-某些工作可以并行","depth":3},{"value":"4. 需要相互质疑，而不是只需要多写一遍","url":"#4-需要相互质疑而不是只需要多写一遍","depth":3},{"value":"三、四种常见协作模式","url":"#三四种常见协作模式","depth":2},{"value":"顺序流水线：一个接一个传递结果","url":"#顺序流水线一个接一个传递结果","depth":3},{"value":"并行专家：多个 Agent 同时给出独立意见","url":"#并行专家多个-agent-同时给出独立意见","depth":3},{"value":"监督者—执行者：负责规划与执行分离","url":"#监督者执行者负责规划与执行分离","depth":3},{"value":"辩论与裁决：不同假设之间互相挑战","url":"#辩论与裁决不同假设之间互相挑战","depth":3},{"value":"四、协作的核心不是聊天，而是共享状态","url":"#四协作的核心不是聊天而是共享状态","depth":2},{"value":"五、消息协议要小，结果要能验证","url":"#五消息协议要小结果要能验证","depth":2},{"value":"六、失败恢复比角色数量更值得设计","url":"#六失败恢复比角色数量更值得设计","depth":2},{"value":"七、如何判断拆分没有成功","url":"#七如何判断拆分没有成功","depth":2},{"value":"八、我的结论：把 Agent 当成有边界的同事","url":"#八我的结论把-agent-当成有边界的同事","depth":2}] Sun, 11 May 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent多智能体架构 https://www.hansking.cn/blog/2026duo-mo-tai-Agent-rang-mo-xing-du-tu-du-biao-du-ping-mu https://www.hansking.cn/blog/2026duo-mo-tai-Agent-rang-mo-xing-du-tu-du-biao-du-ping-mu [{"value":"多模态 Agent：让模型读图、读表、读屏幕","url":"#多模态-agent让模型读图读表读屏幕","depth":1},{"value":"一、先区分三类多模态任务","url":"#一先区分三类多模态任务","depth":2},{"value":"1. 视觉问答","url":"#1-视觉问答","depth":3},{"value":"2. 视觉结构化","url":"#2-视觉结构化","depth":3},{"value":"3. 视觉行动","url":"#3-视觉行动","depth":3},{"value":"二、图像预处理决定模型能不能看清","url":"#二图像预处理决定模型能不能看清","depth":2},{"value":"三、表格不能只当成一张图片","url":"#三表格不能只当成一张图片","depth":2},{"value":"四、结构化输出要有证据和校验","url":"#四结构化输出要有证据和校验","depth":2},{"value":"五、屏幕理解的关键是状态，不是坐标","url":"#五屏幕理解的关键是状态不是坐标","depth":2},{"value":"六、Agent 要知道什么时候自己不确定","url":"#六agent-要知道什么时候自己不确定","depth":2},{"value":"七、评测多模态 Agent 要测整条链路","url":"#七评测多模态-agent-要测整条链路","depth":2},{"value":"八、我最后留下的工程原则","url":"#八我最后留下的工程原则","depth":2}] Tue, 19 May 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM多模态AgentComputer Usetop https://www.hansking.cn/blog/2026da-mo-xing-fa-zhan-shi-zong-fu-pan-cong-sheng-cheng-wen-ben-dao-ke-zhi-xing-xi-tong https://www.hansking.cn/blog/2026da-mo-xing-fa-zhan-shi-zong-fu-pan-cong-sheng-cheng-wen-ben-dao-ke-zhi-xing-xi-tong [{"value":"大模型发展史总复盘：从生成文本到可执行系统","url":"#大模型发展史总复盘从生成文本到可执行系统","depth":1},{"value":"一、2023：模型走出研究论文，进入开发者工具箱","url":"#一2023模型走出研究论文进入开发者工具箱","depth":2},{"value":"二、2024：从“调用模型”走向“组织上下文”","url":"#二2024从调用模型走向组织上下文","depth":2},{"value":"三、2025：模型开始规划、推理和调用工具","url":"#三2025模型开始规划推理和调用工具","depth":2},{"value":"四、2026：真正的重点是系统能不能长期运行","url":"#四2026真正的重点是系统能不能长期运行","depth":2},{"value":"五、四条一直没有改变的主线","url":"#五四条一直没有改变的主线","depth":2},{"value":"1. 数据比想象中更重要","url":"#1-数据比想象中更重要","depth":3},{"value":"2. 评测比感觉更重要","url":"#2-评测比感觉更重要","depth":3},{"value":"3. 边界比能力更重要","url":"#3-边界比能力更重要","depth":3},{"value":"4. 系统比模型更重要","url":"#4-系统比模型更重要","depth":3},{"value":"六、我会怎样继续学习大模型","url":"#六我会怎样继续学习大模型","depth":2},{"value":"七、100 篇之后，真正值得留下的东西","url":"#七100-篇之后真正值得留下的东西","depth":2},{"value":"总结：从“会生成”到“值得托付”","url":"#总结从会生成到值得托付","depth":2}] Sat, 08 Aug 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM技术总结Agent学习路线 https://www.hansking.cn/blog/2026shi-shi-yu-yin-Agent-yan-chi-yu-suan-yu-da-duan-chu-li https://www.hansking.cn/blog/2026shi-shi-yu-yin-Agent-yan-chi-yu-suan-yu-da-duan-chu-li [{"value":"实时语音 Agent：延迟预算与打断处理","url":"#实时语音-agent延迟预算与打断处理","depth":1},{"value":"一、先画出完整音频链路","url":"#一先画出完整音频链路","depth":2},{"value":"二、用延迟预算替代一句“要快一点”","url":"#二用延迟预算替代一句要快一点","depth":2},{"value":"三、VAD 决定系统什么时候接话","url":"#三vad-决定系统什么时候接话","depth":2},{"value":"四、ASR 的 partial 结果不能直接当最终事实","url":"#四asr-的-partial-结果不能直接当最终事实","depth":2},{"value":"五、LLM 和 TTS 要按可播放单元流式协作","url":"#五llm-和-tts-要按可播放单元流式协作","depth":2},{"value":"六、打断是核心能力，不是异常分支","url":"#六打断是核心能力不是异常分支","depth":2},{"value":"七、状态机比一堆布尔变量可靠","url":"#七状态机比一堆布尔变量可靠","depth":2},{"value":"八、实时语音也要有隐私边界","url":"#八实时语音也要有隐私边界","depth":2},{"value":"总结：快不是催出来的，是每一层都愿意让路","url":"#总结快不是催出来的是每一层都愿意让路","depth":2}] Thu, 28 May 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM语音交互AgentWebRTC https://www.hansking.cn/blog/2026chang-ren-wu-Agent-jian-cha-dian-hui-fu-yu-mi-deng-zhi-xing https://www.hansking.cn/blog/2026chang-ren-wu-Agent-jian-cha-dian-hui-fu-yu-mi-deng-zhi-xing [{"value":"长任务 Agent：检查点、恢复与幂等执行","url":"#长任务-agent检查点恢复与幂等执行","depth":1},{"value":"一、先把任务从对话里拿出来","url":"#一先把任务从对话里拿出来","depth":2},{"value":"二、用状态机表达任务生命周期","url":"#二用状态机表达任务生命周期","depth":2},{"value":"三、检查点要保存什么","url":"#三检查点要保存什么","depth":2},{"value":"四、最危险的是“成功了，但我没收到成功响应”","url":"#四最危险的是成功了但我没收到成功响应","depth":2},{"value":"五、重试不是遇到错误就再来一次","url":"#五重试不是遇到错误就再来一次","depth":2},{"value":"六、恢复时不要盲目重跑整个计划","url":"#六恢复时不要盲目重跑整个计划","depth":2},{"value":"七、进度要让用户看得懂","url":"#七进度要让用户看得懂","depth":2},{"value":"八、检查点本身也要版本化和保护","url":"#八检查点本身也要版本化和保护","depth":2},{"value":"九、可靠性评测要模拟真实故障","url":"#九可靠性评测要模拟真实故障","depth":2},{"value":"总结：长任务 Agent 的记忆应该属于系统","url":"#总结长任务-agent-的记忆应该属于系统","depth":2}] Sat, 06 Jun 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent工作流引擎可靠性