
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>XiaoLeiJun</title>
      <link>https://www.xiaoleijun.com/blog</link>
      <description>Focused on implementing low-cost, localized, high-performance, and practical AI applications.</description>
      <language>en-us</language>
      <managingEditor>cuitinglei@qq.com (xiaoleijun)</managingEditor>
      <webMaster>cuitinglei@qq.com (xiaoleijun)</webMaster>
      <lastBuildDate>Thu, 18 Jun 2026 00:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.xiaoleijun.com/tags/agent/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_01</guid>
    <title>Agent Harness 工程：开篇</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_01</link>
    <description>Agent 的核心不只是大模型，而是模型之外那套让它观察环境、调用工具、获取知识、执行动作并接受权限约束的 Harness。本文是 Agent Harness 工程系列的开篇。</description>
    <pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_02</guid>
    <title>Agent Harness 工程：从一次 LLM 调用到 Agent Loop</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_02</link>
    <description>从最基础的一次 ChatCompletion 调用开始，逐步加入工具定义、工具执行和工具结果回传，最后得到一个最小可运行的 Agent Loop。</description>
    <pubDate>Mon, 22 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_03</guid>
    <title>Agent Harness 工程：多工具调用与工具注册表</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_03</link>
    <description>在上一篇最小 Agent Loop 的基础上，继续扩展多个工具，并用工具注册表替代大段 if/elif，让 Harness 能按工具名统一分发、校验和执行工具调用。</description>
    <pubDate>Tue, 23 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_04</guid>
    <title>Agent Harness 工程：权限与审批</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_04</link>
    <description>在多工具 Agent 的基础上加入权限策略层，把工具调用分成直接放行、需要用户审批和直接拒绝三类，并让权限结果也回到 Agent Loop 的消息历史里。</description>
    <pubDate>Wed, 24 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_05</guid>
    <title>Agent Harness 工程：Hook 机制与可扩展循环</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_05</link>
    <description>在权限与审批的基础上继续抽象，把权限检查、日志记录、审计、结果截断等横切逻辑放进 Hook，让 Agent Loop 保持稳定。</description>
    <pubDate>Thu, 25 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_06</guid>
    <title>Agent Harness 工程：TodoWrite 与任务计划</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_06</link>
    <description>在 Hook 机制之后，为 Agent 增加 TodoWrite 工具，让它先拆解任务、持续更新进度，并根据当前 Todo 推进多步骤任务。</description>
    <pubDate>Fri, 26 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_07</guid>
    <title>Agent Harness 工程：Subagent 与上下文隔离</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_07</link>
    <description>在 TodoWrite 之后，为 Agent Harness 增加 Subagent 能力，让主 Agent 负责规划和整合，把局部任务交给干净上下文里的子代理完成。</description>
    <pubDate>Mon, 29 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_08</guid>
    <title>Agent Harness 工程：Skill 与按需加载</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_08</link>
    <description>在 Subagent 之后，继续拆解 Skill 的本质：把任务相关的规范、流程、背景和示例做成可发现、可按需加载的上下文。</description>
    <pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_09</guid>
    <title>Agent Harness 工程：上下文压缩</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_09</link>
    <description>在 Skill 按需加载之后，继续解决 messages 无限膨胀的问题：如何压缩工具结果、摘要历史、保留关键事实，并区分上下文和记忆。</description>
    <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_10</guid>
    <title>Agent Harness 工程：记忆与 Memory Store</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_10</link>
    <description>在上下文压缩之后，继续拆解 Agent 记忆系统：哪些信息应该长期保存，Memory Store 如何设计，以及如何避免记住错误事实和敏感信息。</description>
    <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_11</guid>
    <title>Agent Harness 工程：提示词的运行时组装</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_11</link>
    <description>在记忆系统之后，继续拆解提示词的运行时组装：如何把固定 SYSTEM_PROMPT 拆成可组合的 Section，并根据当前运行态按需加载。</description>
    <pubDate>Tue, 07 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_12</guid>
    <title>Agent Harness 工程：错误恢复与重试</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_12</link>
    <description>在运行时提示词组装之后，继续为 Agent 增加错误恢复能力：识别输出截断、上下文超限和临时故障，并为不同失败选择有限、可观测的恢复策略。</description>
    <pubDate>Wed, 08 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_13</guid>
    <title>Agent Harness 工程：Task System 与持久化任务图</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_13</link>
    <description>在错误恢复之后，把一次性的 Todo 升级为有依赖、有状态、可持久化的任务图，并通过任务认领、租约和 checkpoint 支持跨会话恢复。</description>
    <pubDate>Thu, 09 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_14</guid>
    <title>Agent Harness 工程：后台任务与异步通知</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_14</link>
    <description>在持久化任务图之后，为 Agent 增加后台执行能力：启动耗时命令后立即返回 job_id，并通过状态轮询、事件通知和输出工具继续推进其它任务。</description>
    <pubDate>Mon, 13 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.xiaoleijun.com/blog/agent_learn_15</guid>
    <title>Agent Harness 工程：定时任务与 Scheduler</title>
    <link>https://www.xiaoleijun.com/blog/agent_learn_15</link>
    <description>在后台 Job 之后，为 Agent 增加持久化 Scheduler：按 Cron 与时区触发工作，并通过 ScheduleRun、Task System、错过执行和并发策略保证定时任务可恢复、可追踪。</description>
    <pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate>
    <author>cuitinglei@qq.com (xiaoleijun)</author>
    <category>AI</category><category>Agent</category><category>Harness</category><category>工程实践</category>
  </item>

    </channel>
  </rss>
