<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>孙笑笑 · likely</title>
    <link>https://tntlikely-home.pages.dev</link>
    <description>全栈工程师 &amp; 开源作者的技术调研与写作</description>
    <language>zh-CN</language>
    <atom:link href="https://tntlikely-home.pages.dev/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Agent 服务的并发与状态：异步、任务队列、幂等和故障恢复怎么配合</title>
      <link>https://tntlikely-home.pages.dev/articles/agent-concurrency-state</link>
      <guid>https://tntlikely-home.pages.dev/articles/agent-concurrency-state</guid>
      <pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate>
      <description>把厂商会话能力与业务一致性分开，解释 async I/O、有限并发、事务 outbox、任务租约和稳定操作 ID；用崩溃窗口、慢客户端与外部结果未知来检验生产服务。</description>
      <category>Agent</category>
      <category>服务端</category>
      <category>并发</category>
      <category>一致性</category>
    </item>
    <item>
      <title>Agent 怎么持续优化与上线：Trace、评测、SLO 和灰度交付的证据链</title>
      <link>https://tntlikely-home.pages.dev/articles/agent-observability-delivery</link>
      <guid>https://tntlikely-home.pages.dev/articles/agent-observability-delivery</guid>
      <pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate>
      <description>对照 OpenAI tracing 与 harness 实践、Anthropic 长任务和评测方法，建立跨模型、工具、队列的观测链；解释 SLO、费用归因、Docker/K8s 发布和可回滚的优化实验。</description>
      <category>Agent</category>
      <category>可观测性</category>
      <category>评测</category>
      <category>CI/CD</category>
    </item>
    <item>
      <title>生产级 Agent 怎么设计：从 OpenAI 与 Anthropic 的公开实践看架构边界</title>
      <link>https://tntlikely-home.pages.dev/articles/production-agent-design</link>
      <guid>https://tntlikely-home.pages.dev/articles/production-agent-design</guid>
      <pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate>
      <description>对照 Codex agent loop、Agents SDK/API 与 Claude Managed Agents，解释模型、执行框架、会话和业务状态的职责；从成功标准出发，设计可恢复执行、权限边界与优化实验。</description>
      <category>Agent</category>
      <category>OpenAI</category>
      <category>Anthropic</category>
      <category>生产架构</category>
    </item>
    <item>
      <title>Prompt 与上下文工程：OpenAI、Anthropic 的方法，版本发布与优化实验</title>
      <link>https://tntlikely-home.pages.dev/articles/prompt-context-engineering</link>
      <guid>https://tntlikely-home.pages.dev/articles/prompt-context-engineering</guid>
      <pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate>
      <description>从官方提示实践与长上下文研究出发，解释指令、证据、历史和工具如何组合；用版本化发布包、受控实验、影子验证与灰度回滚，把 prompt 优化纳入工程流程。</description>
      <category>Prompt</category>
      <category>上下文工程</category>
      <category>OpenAI</category>
      <category>Anthropic</category>
    </item>
    <item>
      <title>Vibe Coding 的工程纪律：任务、审查与验证怎么分工</title>
      <link>https://tntlikely-home.pages.dev/articles/vibe-coding-discipline</link>
      <guid>https://tntlikely-home.pages.dev/articles/vibe-coding-discipline</guid>
      <pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate>
      <description>对照 OpenAI harness engineering、Claude Code 与生产力研究，解释任务契约、仓库知识、机械约束和独立验收；把审查、返工与实际交付一起计入优化。</description>
      <category>AI 编程</category>
      <category>Claude Code</category>
      <category>工程效率</category>
      <category>Vibe Coding</category>
    </item>
    <item>
      <title>面向未成年人的 AI 产品：合规要求与安全工程</title>
      <link>https://tntlikely-home.pages.dev/articles/minors-ai-safety-engineering</link>
      <guid>https://tntlikely-home.pages.dev/articles/minors-ai-safety-engineering</guid>
      <pubDate>Tue, 15 Sep 2026 00:00:00 GMT</pubDate>
      <description>核对法规适用范围与年龄边界，补齐生成内容标识、授权和删除链路；把数据保护、流式审核、风险分流、防依赖与人工处置落实到工程。</description>
      <category>安全合规</category>
      <category>未成年人保护</category>
      <category>内容安全</category>
      <category>LLM</category>
    </item>
    <item>
      <title>端云协同 API 设计：离线同步、弱网重试与版本兼容</title>
      <link>https://tntlikely-home.pages.dev/articles/device-cloud-api-sync</link>
      <guid>https://tntlikely-home.pages.dev/articles/device-cloud-api-sync</guid>
      <pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate>
      <description>用本地事务与 outbox 保住离线记录，以业务操作 ID 去重；设计逐项确认、游标失效、删除传播、冲突策略和内容包原子切换，并验证断电与重连。</description>
      <category>架构</category>
      <category>API 设计</category>
      <category>离线优先</category>
      <category>端云协同</category>
    </item>
    <item>
      <title>评测闭环：让每次 Prompt 变更都可回归</title>
      <link>https://tntlikely-home.pages.dev/articles/llm-eval-loop</link>
      <guid>https://tntlikely-home.pages.dev/articles/llm-eval-loop</guid>
      <pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate>
      <description>从小规模回归集起步，区分确定性校验、模型裁判与人工评审；给出技术问答 rubric、统计边界、版本记录和上线门槛，避免用平均分掩盖退化。</description>
      <category>LLM</category>
      <category>评测</category>
      <category>Evals</category>
      <category>质量</category>
    </item>
    <item>
      <title>多 Agent 如何通信：任务书、结构化回传与 A2A 协议</title>
      <link>https://tntlikely-home.pages.dev/articles/multi-agent-communication</link>
      <guid>https://tntlikely-home.pages.dev/articles/multi-agent-communication</guid>
      <pubDate>Tue, 16 Jun 2026 00:00:00 GMT</pubDate>
      <description>从可并行任务出发，设计输入、状态、成果与失败契约；说明共享文件和跨机引用的区别、MCP 与 A2A 的职责，以及协调成本与验收方法。</description>
      <category>Agent</category>
      <category>多智能体</category>
      <category>架构</category>
      <category>A2A</category>
      <category>MCP</category>
    </item>
    <item>
      <title>Claude Code 本地 Agent 的安全栈：沙盒、权限、规则与分类器</title>
      <link>https://tntlikely-home.pages.dev/articles/claude-code-local-agent-security</link>
      <guid>https://tntlikely-home.pages.dev/articles/claude-code-local-agent-security</guid>
      <pubDate>Tue, 19 May 2026 00:00:00 GMT</pubDate>
      <description>区分提示规则、程序权限与 OS 沙盒，核对 Claude Code 的真实覆盖范围和平台差异；解释凭证掩码、非沙盒组件和可重复的边界验证。</description>
      <category>Claude Code</category>
      <category>Agent</category>
      <category>安全</category>
      <category>沙盒</category>
    </item>
    <item>
      <title>上下文与记忆管理：Agent 的长期记忆怎么建</title>
      <link>https://tntlikely-home.pages.dev/articles/context-memory-management</link>
      <guid>https://tntlikely-home.pages.dev/articles/context-memory-management</guid>
      <pubDate>Tue, 21 Apr 2026 00:00:00 GMT</pubDate>
      <description>区分会话状态、历史压缩与跨会话记忆；用来源、作用域、版本和有效期治理记忆，解释读写流程、遗忘、权限隔离与可验证的评测。</description>
      <category>Agent</category>
      <category>上下文管理</category>
      <category>记忆</category>
      <category>LLM</category>
    </item>
    <item>
      <title>Function Calling 与结构化输出：让 LLM 可靠地动手</title>
      <link>https://tntlikely-home.pages.dev/articles/function-calling-structured-output</link>
      <guid>https://tntlikely-home.pages.dev/articles/function-calling-structured-output</guid>
      <pubDate>Tue, 17 Mar 2026 00:00:00 GMT</pubDate>
      <description>分清调用协议与输出格式，比较 OpenAI、Claude 的字段；用完整 schema 和工具循环说明参数验证、授权、并发、幂等与失败处理。</description>
      <category>LLM</category>
      <category>Tool Use</category>
      <category>Agent</category>
      <category>结构化输出</category>
    </item>
    <item>
      <title>RAG 全链路（三）：重排、引用溯源与效果评估</title>
      <link>https://tntlikely-home.pages.dev/articles/rag-03-rerank-citation-eval</link>
      <guid>https://tntlikely-home.pages.dev/articles/rag-03-rerank-citation-eval</guid>
      <pubDate>Tue, 17 Feb 2026 00:00:00 GMT</pubDate>
      <description>按 token 预算选择重排后的证据，区分引用存在、引用支撑与事实正确；分层评估检索、生成和业务结果，并处理资料冲突、缺证据与超时。</description>
      <category>RAG</category>
      <category>Rerank</category>
      <category>评测</category>
      <category>可信</category>
    </item>
    <item>
      <title>RAG 全链路（二）：向量库选型与混合检索</title>
      <link>https://tntlikely-home.pages.dev/articles/rag-02-vector-hybrid-search</link>
      <guid>https://tntlikely-home.pages.dev/articles/rag-02-vector-hybrid-search</guid>
      <pubDate>Tue, 20 Jan 2026 00:00:00 GMT</pubDate>
      <description>按查询、权限过滤和运维条件选择向量库；解释向量与词法检索的互补、RRF 融合，以及 ANN 召回和业务证据召回的区别。</description>
      <category>RAG</category>
      <category>向量数据库</category>
      <category>检索</category>
      <category>pgvector</category>
    </item>
    <item>
      <title>RAG 全链路（一）：文档切分——让检索结果保留完整语境</title>
      <link>https://tntlikely-home.pages.dev/articles/rag-01-chunking</link>
      <guid>https://tntlikely-home.pages.dev/articles/rag-01-chunking</guid>
      <pubDate>Tue, 16 Dec 2025 00:00:00 GMT</pubDate>
      <description>从解析和元数据开始，比较固定窗口、结构化切分、父子块与上下文补注；用一份故障手册实例解释检索粒度、供给粒度和版本追踪。</description>
      <category>RAG</category>
      <category>检索</category>
      <category>Embedding</category>
      <category>LLM</category>
    </item>
    <item>
      <title>把失败当默认前提：LLM 服务的重试、降级与兜底设计</title>
      <link>https://tntlikely-home.pages.dev/articles/llm-fallback-design</link>
      <guid>https://tntlikely-home.pages.dev/articles/llm-fallback-design</guid>
      <pubDate>Tue, 18 Nov 2025 00:00:00 GMT</pubDate>
      <description>按错误语义决定重试，用总截止时间控制等待；分清模型重复调用与业务幂等，处理流式中断、缓存串用和降级质量，并用 SLO 验证兜底。</description>
      <category>LLM</category>
      <category>可靠性</category>
      <category>架构</category>
      <category>SLA</category>
    </item>
    <item>
      <title>首 token 为什么这么慢：TTFT 的构成与优化清单</title>
      <link>https://tntlikely-home.pages.dev/articles/ttft-first-token-latency</link>
      <guid>https://tntlikely-home.pages.dev/articles/ttft-first-token-latency</guid>
      <pubDate>Tue, 21 Oct 2025 00:00:00 GMT</pubDate>
      <description>分清首事件、首 token 与首字上屏，拆解检索、排队、prefill 和渲染的延迟；用分阶段埋点、缓存对照与峰值压测找到真正瓶颈。</description>
      <category>LLM</category>
      <category>推理</category>
      <category>延迟优化</category>
      <category>流式</category>
    </item>
    <item>
      <title>控制 LLM 预算：Token 计数、用量统计与没有官方数据时的做法</title>
      <link>https://tntlikely-home.pages.dev/articles/llm-budget-token-accounting</link>
      <guid>https://tntlikely-home.pages.dev/articles/llm-budget-token-accounting</guid>
      <pubDate>Tue, 16 Sep 2025 00:00:00 GMT</pubDate>
      <description>分清请求前预估、响应 usage 和最终账单，避免缓存与推理 token 重复计费；用预算预占、逐次结算和任务成本指标管住长会话。</description>
      <category>LLM</category>
      <category>成本工程</category>
      <category>Agent</category>
      <category>Claude Code</category>
      <category>Codex</category>
    </item>
  </channel>
</rss>