<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>OPALL · 判断与档案</title>
    <link>https://opallagent.com/</link>
    <atom:link href="https://opallagent.com/feed.xml" rel="self" type="application/rss+xml"/>
    <description>OPALL Field Notes 与研习底稿:围绕 AI 与 Agent 的判断、边界、证据链。研习学基础,笔记读判断。</description>
    <language>zh-cn</language>
    <lastBuildDate>Sun, 05 Jul 2026 14:00:00 GMT</lastBuildDate>
    <item>
      <title>P-007 · Signal Pipeline：单人内容信号管道开源模板</title>
      <link>https://opallagent.com/projects/signal-pipeline.html</link>
      <guid>https://opallagent.com/projects/signal-pipeline.html</guid>
      <pubDate>Sat, 04 Jul 2026 11:00:00 GMT</pubDate>
      <description>每日采集公开信号 → 模板化预筛 → 人工审核 → 四个内容出口；零依赖、逐日台账、Actions 云端采集。发布必须人工。</description>
    </item>
    <item>
      <title>P-006 · Minimal Agent Loop：零依赖的最小 Agent 闭环教学库</title>
      <link>https://opallagent.com/projects/minimal-agent-loop.html</link>
      <guid>https://opallagent.com/projects/minimal-agent-loop.html</guid>
      <pubDate>Sat, 04 Jul 2026 10:30:00 GMT</pubDate>
      <description>循环、工具权限边界、trace、重试熔断、独立验证——克隆即跑，离线模式不需要任何 API key。</description>
    </item>
    <item>
      <title>P-003 · OPALL Knowledge System：内容治理方法与证据边界</title>
      <link>https://opallagent.com/projects/opall-knowledge-system.html</link>
      <guid>https://opallagent.com/projects/opall-knowledge-system.html</guid>
      <pubDate>Sat, 04 Jul 2026 10:00:00 GMT</pubDate>
      <description>把私有资料、项目复盘、联网校验、内容分级和证据索引收敛为 public-safe 公开表达的治理项目。</description>
    </item>
    <item>
      <title>P-002 · Business Agent Workflow：业务 Agent 工作流脱敏概览</title>
      <link>https://opallagent.com/projects/business-agent-workflow.html</link>
      <guid>https://opallagent.com/projects/business-agent-workflow.html</guid>
      <pubDate>Sat, 04 Jul 2026 09:30:00 GMT</pubDate>
      <description>面向邮件、报价、知识库和任务流的工作流概览：只保留抽象流程、人工审核点和证据留存方式。</description>
    </item>
    <item>
      <title>P-001 · OpenAgent Evaluation Platform：可追问的 Coding Agent 评测证据框架</title>
      <link>https://opallagent.com/projects/openagent.html</link>
      <guid>https://opallagent.com/projects/openagent.html</guid>
      <pubDate>Sat, 04 Jul 2026 09:00:00 GMT</pubDate>
      <description>围绕 TaskSpec、workspace isolation、trace、quality gate、replay 与 scorecard 的评测证据框架。</description>
    </item>
    <item>
      <title>N-015 · 沙箱即权限工程：怎么评一个隔离方案</title>
      <link>https://opallagent.com/notes/sandbox-as-permission.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/sandbox-as-permission.html</guid>
      <description>核心判断：沙箱不是安全插件，是权限工程——评隔离方案评的是边界定义、生命周期、出入管制、可审计性四个维度。判断对象：采购与自建隔离方案时怎么问对问题。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-014 · 多 Agent 的协作边界：什么时候多反而更糟</title>
      <link>https://opallagent.com/notes/multi-agent-boundary.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/multi-agent-boundary.html</guid>
      <description>核心判断：加第二个 Agent 的理由必须指向结构——职责分离、上下文分工、权限隔离；多 Agent 放大的是结构，不是质量。判断对象："以协作换质量"的常见误用。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-013 · 当 Agent 遇到企业围墙：禁、隔离、审计的三段路</title>
      <link>https://opallagent.com/notes/agent-enterprise-walls.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/agent-enterprise-walls.html</guid>
      <description>核心判断：企业对 Agent 工具的反应不是接受或拒绝二选一，而是禁→隔离→审计的三段路；禁不是终点，是议价的起点。判断对象：Agent 工具进入企业的信任工程化路径。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-012 · AI 时代的技术面试，考察的东西变了</title>
      <link>https://opallagent.com/notes/ai-era-interviews.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/ai-era-interviews.html</guid>
      <description>核心判断：当产出变得便宜，考产出的面试就失效——考判断的面试进场，追问与亲历成为新的可信信号。判断对象：技术面试的换轴与准备重心（趋势自公开报道，亲历已脱敏）。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-011 · AI 不是下一个工具浪潮，而是责任结构的重写</title>
      <link>https://opallagent.com/notes/responsibility-rewrite.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/responsibility-rewrite.html</guid>
      <description>核心判断：AI 带来的组织变化主要不是效率变化，而是签字、审核、留痕三件事的重写。判断对象：组织与个人在新责任结构中的位置（阅读路径第一层）。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-010 · 没有实验室和算力，如何做出可追问的 Agent 项目</title>
      <link>https://opallagent.com/notes/agent-project-without-lab.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/agent-project-without-lab.html</guid>
      <description>核心判断：Agent 工程项目的门槛不是算力和实验室，而是闭环与证据。判断对象：在校生与转方向工程师的项目路线与证据链。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-009 · 长任务生成系统的工程边界</title>
      <link>https://opallagent.com/notes/long-task-pipeline-boundary.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/long-task-pipeline-boundary.html</guid>
      <description>核心判断：生成式 3D 等长任务系统的可信度不来自单次生成的惊艳，而来自任务状态、失败处理、质量门和成本结构。判断对象：公开可观察的生成式 3D 产品与 API，仅作为行业公开样本，不含任何内部信息。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-008 · 业务流程如何转成 Agent 审核闭环</title>
      <link>https://opallagent.com/notes/business-process-to-agent-loop.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/business-process-to-agent-loop.html</guid>
      <description>核心判断：接入 Agent 不是替换某个人，而是把流程拆成建议、审核、动作、反馈四类可复盘状态。判断对象：业务流程引入 Agent 的正确切入方式。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-007 · 为什么 Coding Agent 评测需要 trace 和 replay</title>
      <link>https://opallagent.com/notes/coding-agent-eval-trace.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/coding-agent-eval-trace.html</guid>
      <description>核心判断：通过率数字会说谎；任务定义、环境边界、trace、产物和验证结果都被保留，评测才能被审计。判断对象：Coding Agent 评测系统的可信度。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-006 · 边界不清的 Agent 项目为什么容易失败</title>
      <link>https://opallagent.com/notes/unbounded-agent-failure.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/unbounded-agent-failure.html</guid>
      <description>核心判断：Agent 项目失败的主因不是模型不够强，而是权限、异常、审核、日志和回退被省略。判断对象：从演示走向生产的 Agent 项目的边界设计。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-005 · 高端 AI 项目为什么先看证据链</title>
      <link>https://opallagent.com/notes/evidence-chain-first.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/evidence-chain-first.html</guid>
      <description>核心判断：README、trace、artifact、测试和边界声明，比顺滑 demo 更能支撑判断。判断对象：评估高端 AI 项目时的证据链体检方法。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-004 · 报价 Agent 的真实边界</title>
      <link>https://opallagent.com/notes/quotation-agent-boundary.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/quotation-agent-boundary.html</guid>
      <description>核心判断：报价 Agent 的价值不是自动报价，而是发现缺失条件、整理假设、提示风险，把价格承诺交给人工确认。判断对象：AI 进入报价与销售承诺流程时的确认权与责任边界。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-003 · AI 项目如何变成可追问证据</title>
      <link>https://opallagent.com/notes/project-evidence.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/project-evidence.html</guid>
      <description>核心判断：一个 AI 项目要变成可追问证据，至少需要三类材料——Artifact 说明做了什么，Boundary 说明没做什么，Review 说明别人可以从哪里追问。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-002 · 为什么私有知识系统应该先于 RAG</title>
      <link>https://opallagent.com/notes/private-knowledge-before-rag.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/private-knowledge-before-rag.html</guid>
      <description>核心判断：RAG 不是知识系统的第一步。先稳定来源、权限、版本、证据结构和人工审核流程。判断对象：内部知识源、项目文档或业务知识进入 AI 系统前的来源、权限、版本和证据边界。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>N-001 · 为什么业务 Agent 不应一开始自动发邮件</title>
      <link>https://opallagent.com/notes/business-agent-email-boundary.html</link>
      <guid isPermaLink="true">https://opallagent.com/notes/business-agent-email-boundary.html</guid>
      <description>核心判断：第一版业务 Agent 应辅助判断、生成草稿并保留人工审核，而不是直接替人发送邮件。判断对象：AI 接入邮件、销售、报价或客户沟通流程时的发送权、审核权和责任边界。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-015 · 第一个 AI 项目怎么选题</title>
      <link>https://opallagent.com/study/first-project.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/first-project.html</guid>
      <description>一句话：好的第一题，三样都离你近：近的摩擦、近的闭环、近的验证。宁可题小闭环全，不要题大闭环断。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-014 · 技能包：Agent 的能力开始像软件包</title>
      <link>https://opallagent.com/study/agent-skills.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/agent-skills.html</guid>
      <description>一句话：能力从"接线"变成"装包"：可分发、可安装、可复用。装包前读说明书、最小安装、看权限——软件包的老教训一条不落。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-013 · 怎么知道你的 Agent 变好了：评测入门</title>
      <link>https://opallagent.com/study/agent-evals.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/agent-evals.html</guid>
      <description>一句话："感觉变好了"多半靠不住。一张不换的考卷：固定的题、固定的判卷、可重跑——题库比手感诚实。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-012 · Agent 的房间：沙箱与权限入门</title>
      <link>https://opallagent.com/study/agent-sandbox.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/agent-sandbox.html</guid>
      <description>一句话：对 Agent 的信任，不该建立在"它不会犯错"上，该建立在"犯错也出不了这间房"上。三面墙管过程，一扇门管结果。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-011 · 上下文工程：提示词之后的新基本功</title>
      <link>https://opallagent.com/study/context-engineering.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/context-engineering.html</guid>
      <description>一句话：上下文不是聊天记录，是每一步都要重新布置的工作台。长任务变笨，先看台面上四类东西哪类腐坏了。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-010 · 作品集里的 AI 项目怎么经得起追问</title>
      <link>https://opallagent.com/study/portfolio-under-questioning.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/portfolio-under-questioning.html</guid>
      <description>一句话：作品集的单位不是项目数量，是能接住的追问数量。四样东西反过来留全，发出去前先过两句追问测试。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-009 · 用 AI 学，还是被 AI 代学：界线在哪</title>
      <link>https://opallagent.com/study/learn-or-delegate.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/learn-or-delegate.html</guid>
      <description>一句话：用 AI 学和被 AI 代学，当下手感一模一样。检验只有一句：离开 AI，还能复述吗——界线不在用量，在谁做决策。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-008 · 从抄到改到造：教程看了很多，项目为什么还是做不出来</title>
      <link>https://opallagent.com/study/from-copy-to-create.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/from-copy-to-create.html</guid>
      <description>一句话：教程只给结果，项目需要决策。中间缺的那级台阶叫"改"——崩的位置，就是决策所在的位置。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-007 · 多 Agent 协作的基本形态</title>
      <link>https://opallagent.com/study/multi-agent-patterns.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/multi-agent-patterns.html</guid>
      <description>一句话：多 Agent 不是更强的 Agent，是组织结构的选择。流水线、评审团、分诊台——什么问题配什么结构。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-006 · MCP 与工具调用：AI 怎么伸出手</title>
      <link>https://opallagent.com/study/mcp-and-tool-calling.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/mcp-and-tool-calling.html</guid>
      <description>一句话：模型只会说话，动作全是外面装上去的。函数调用的四步握手、MCP 统一插座，和三层缰绳——能力的来源和边界。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-005 · 提示词、微调与工具：三种改变模型行为的方式</title>
      <link>https://opallagent.com/study/prompt-finetune-tools.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/prompt-finetune-tools.html</guid>
      <description>一句话：便签、参考书、培训——顺序永远是提示词 → 增强 → 微调；不对味修提示词，不知道上增强，不稳定才谈微调。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-004 · 上下文、token 与成本的直觉</title>
      <link>https://opallagent.com/study/context-tokens-cost.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/context-tokens-cost.html</guid>
      <description>一句话：模型没有记忆，只有重读——这一条解释了长对话为什么变贵变慢变笨，也解释了缓存省在哪里。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-003 · 怎么读一个 AI 项目：评测与证据的入门</title>
      <link>https://opallagent.com/study/how-to-read-an-ai-project.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/how-to-read-an-ai-project.html</guid>
      <description>一句话：看 AI 项目先找四样东西——任务定义、运行轨迹、验证方式、边界声明。四样俱全是工程，四样全无是表演。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-002 · 私有知识与 RAG 的部件清单</title>
      <link>https://opallagent.com/study/rag-parts-list.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/rag-parts-list.html</guid>
      <description>一句话：RAG 是五个部件的流水线，每个部件都有自己的失败方式。会拆部件，才知道"检索不准"到底该修哪里。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
    <item>
      <title>F-001 · Agent 到底是什么：从对话到动作的四级阶梯</title>
      <link>https://opallagent.com/study/what-is-an-agent.html</link>
      <guid isPermaLink="true">https://opallagent.com/study/what-is-an-agent.html</guid>
      <description>一句话：Agent 不是更聪明的聊天，而是被授权采取动作的循环。理解它的关键不是模型，而是"授权"两个字。</description>
      <pubDate>Sat, 04 Jul 2026 12:00:00 GMT</pubDate>
    </item>
  </channel>
</rss>
