AI Trending
All 全部 Deep Analysis 深度解析 Foresight 前瞻 AI News AI资讯 Open Source 开源项目 AI Products AI产品 Research Papers 论文研究 AI Security AI安全 AI Practices AI实践 AI Skills AI技能 AI Overseas AI出海
AI Skills AI技能 8h ago • Updated 1h ago 更新于 1小时前 43

LAI #142: My AI Setup in 2026 LAI #142：我的2026年AI工作流

The author shares a portable AI workflow using Claude Code for planning/coordination and Codex for bounded execution, with knowledge stored in an open-source Obsidian vault accessible across devices "AI Engineering for Production" (launching October 20) addresses engineering challenges that better models alone cannot solve: context management, retrieval, agents, evaluation, recovery, and deployment Document parsing quality is a critical but underestimated bottleneck—extraction can break structur 作者分享2026年个人AI工作流：使用Obsidian vault管理可移植知识与技能，通过Mac mini运行Claude Code（规划协调）和Codex（执行任务），实现跨设备访问 新书《AI Engineering for Production》将于10月20日发布，聚焦模型能力之外的工程问题：context、retrieval、agents、evaluation、recovery、deployment 多篇技术文章深入探讨：Continuous Batching提升GPU利用率、多Agent共享内存的并发问题（丢失更新/脏读/幻读）、Agent沙箱的动态网络权限策略、以及基于上下文反

62
Hot 热度
68
Quality 质量
55
Impact 影响力

Analysis 深度分析

TL;DR

  • The author shares a portable AI workflow using Claude Code for planning/coordination and Codex for bounded execution, with knowledge stored in an open-source Obsidian vault accessible across devices
  • "AI Engineering for Production" (launching October 20) addresses engineering challenges that better models alone cannot solve: context management, retrieval, agents, evaluation, recovery, and deployment
  • Document parsing quality is a critical but underestimated bottleneck—extraction can break structural relationships that retrieval and prompting cannot recover
  • Agent frameworks leave four operational gaps unaddressed: context selection, observability, scalability, and governance, requiring infrastructure beneath interfaces like MCP
  • Shared-memory multi-agent systems introduce classic database concurrency problems (lost updates, write skew, dirty reads) that consistency mechanisms alone cannot fully resolve

Why It Matters

This article synthesizes practical workflow insights with deeper engineering principles relevant to anyone building production AI systems. The emphasis on portable knowledge, operational gaps in agent frameworks, and concurrency challenges in shared-memory agents directly addresses the gap between demo-quality systems and reliable production deployments.

Technical Details

  • Portable Agent Workflow: Claude Code handles planning and coordination; Codex handles bounded execution and scheduled tasks. Knowledge and skills are maintained in an Obsidian vault, making them device-agnostic and model-agnostic.
  • Document Parsing Pitfalls: PDF extraction can lose structural relationships—table values detached from column headings, footnotes displaced from qualifying sentences. The recommended debugging test is to attempt answering an incorrect system response using only the extracted text.
  • Continuous Batching: Replaces static batching by removing completed sequences and admitting new requests at every generation step, improving GPU utilization. Performance should be judged across utilization, throughput, latency, and queueing behavior rather than a single speedup metric.
  • Agent Framework Operational Gaps: Four areas remain unaddressed by frameworks: context selection, observability, scalability, and governance. Durable state, retrieval, recovery, permissions, and monitoring require infrastructure beneath interfaces like Store and Checkpointer.
  • Shared-State Concurrency in Multi-Agent Systems: Lost updates, write skew, dirty reads, phantom reads, and stale-context writes arise when agents concurrently modify shared memory. Solutions include partitioning, single-writer designs, reducers, event logs, optimistic concurrency control, and CRDTs—but semantic reconciliation and idempotency remain application-layer concerns.
  • Sandbox Egress Policies: Network access should be phase-dependent rather than fixed for the sandbox lifetime. Tensorlake Sandboxes demonstrate dynamic egress policy replacement, restricting access after dependency installation completes.
  • Contextual Off-Policy Evaluation: New policies that perform better on average may harm specific subpopulations. Cross-fitted doubly robust estimates with confidence bounds can gate deployment to contexts where evidence supports improvement.

Industry Insight

  • The shift from model-centric to engineering-centric thinking is accelerating—books, frameworks, and community discourse now focus on reliability, deployment, and operational concerns rather than raw model capability.
  • Multi-agent systems will increasingly require database-grade concurrency controls and sandbox isolation patterns borrowed from distributed systems engineering.
  • Token usage should be evaluated through outcome quality rather than raw cost, as agents capable of end-to-end task execution (research, implement, test, iterate) can justify higher consumption when aligned with productive work.

TL;DR

  • 作者分享2026年个人AI工作流：使用Obsidian vault管理可移植知识与技能，通过Mac mini运行Claude Code（规划协调）和Codex（执行任务），实现跨设备访问
  • 新书《AI Engineering for Production》将于10月20日发布，聚焦模型能力之外的工程问题：context、retrieval、agents、evaluation、recovery、deployment
  • 多篇技术文章深入探讨：Continuous Batching提升GPU利用率、多Agent共享内存的并发问题（丢失更新/脏读/幻读）、Agent沙箱的动态网络权限策略、以及基于上下文反事实评估的策略部署方法
  • 社区开源项目Alineo提供隔离环境供coding agents安全运行，支持状态保存、快照恢复和并行fork

为什么值得看

本文汇集了当前AI工程实践的核心痛点与解决方案，从个人工作流到生产级架构均有覆盖，对AI从业者的日常实践和系统设计具有直接参考价值。精选的技术文章深入探讨了GPU调度优化、多Agent并发安全、沙箱权限管理等生产环境的关键问题。

技术解析

  • Continuous Batching机制：传统静态batching固定请求组直到所有序列完成，短请求会浪费GPU容量。Continuous batching在每个生成步骤动态移除已完成序列并接纳等待请求，需通过utilization、throughput、latency和queueing behavior综合评估serving系统性能，而非单一加速比。
  • 多Agent共享内存的并发问题：当多个Agent并发读写同一内存时，会出现lost updates、write skew、dirty reads、phantom reads、stale-context writes等数据库级并发问题。解决方案包括partitioning、single-writer设计、reducers、event logs、optimistic concurrency control和CRDTs，但一致性机制无法解决语义冲突或重复执行问题，需应用层reconciliation和idempotency。
  • Agent沙箱动态权限策略：Agent安装依赖时需要广泛网络访问，但执行不可信代码时无需保持该权限。Tensorlake Sandboxes支持按任务阶段动态替换egress policy，而非固定整个sandbox生命周期，需处理allowlists、internet-access设置、策略更新失败和现有连接等问题。
  • 上下文反事实策略评估：新策略平均表现更好但仍可能对部分群体产生负面影响。通过cross-fitted doubly robust estimates结合confidence bounds构建gate，仅在证据支持的上下文中切换策略，避免一刀切部署。
  • Agent框架vs运营平台：框架定义工作流但不解决底层运维问题，缺失层包括context selection、observability、scalability、governance。模型能力和MCP等标准会变化，但durable state、retrieval、recovery、permissions、monitoring需要基础设施支撑。

行业启示

  • AI工程的重心正从"模型能力竞赛"转向"生产可靠性工程"，context管理、retrieval质量、agent并发安全、评估与恢复机制成为差异化竞争关键，从业者需建立系统化的工程思维而非仅关注模型选型。
  • 多Agent系统的内存和状态管理是尚未成熟的基础设施问题，需要借鉴数据库并发控制理论（如CRDTs、event sourcing）并结合应用层语义协调，沙箱安全策略也应从静态权限转向动态阶段化控制。
  • 策略部署和模型路由需精细化：基于上下文的反事实评估可避免"平均提升但局部损害"的陷阱，不同任务类型应路由到不同模型（如Claude Code用于规划、Codex用于执行），token消耗应评估ROI而非单纯追求最低成本。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明：以上内容由 AI 生成，仅供参考。

Claude Claude Agent Agent Code Generation 代码生成 LLM 大模型 GPU GPU
Read Original → 阅读原文 →

Related Articles 相关文章

Anthropic Surpasses OpenAI: The 'Code is King' Logic Behind $965 Billion Valuation Anthropic超越OpenAI：9650亿估值背后的「代码为王」逻辑
Show HN: Security Cards – Reducing insecure AI-generated code by 72% 展示 HN：安全卡片——将不安全 AI 生成代码减少 72%
[GitHub] apache/texera GitHub 上的 Apache Texera 项目
GPT-5.6 vs Claude Opus 4.8 vs MiniMax M3: A Three-Way Battle, Who is Leading? GPT-5.6 vs Claude Opus 4.8 vs MiniMax M3：三强争霸，谁在领跑？
New Deepseek model V4.1-Flash cuts memory needs for AI agents Deepseek发布V4.1-Flash新模型，大幅降低AI代理内存需求