Multi-Agent on Lost Templehttps://cloudcold.ai/tags/multi-agent/Recent content in Multi-Agent on Lost TempleHugoenSun, 07 Jun 2026 00:00:00 +0000Google Agentic RAG：不再差不多的企业级 RAG — Sufficient Context Agent 深度拆解https://cloudcold.ai/posts/2026-06-07-agentic-rag-google-deep-dive/Sun, 07 Jun 2026 00:00:00 +0000https://cloudcold.ai/posts/2026-06-07-agentic-rag-google-deep-dive/Google Gemini Enterprise Agent Platform 的 Agentic RAG 框架深度拆解：Sufficient Context Agent 三重检查机制、五阶段流水线、Cross-Corpus Routing。FramesQA 基准测试准确率提升 34%，跨 4 数据源准确率 90.1%。4-Agent Pipeline: The Seductive Trap of 'Ship While You Sleep'https://cloudcold.ai/posts/2026-06-06-4-agent-pipeline-seductive-trap/Sat, 06 Jun 2026 00:00:00 +0000https://cloudcold.ai/posts/2026-06-06-4-agent-pipeline-seductive-trap/四个 AI Agent 的 Planner-Coder-Tester-Reviewer 流水线真正擅长的是小任务。Token 成本 4-7x 基线、无 feedback loop、silent failure 系统性存在。Anthropic 官方立场直接矛盾。学术背刺：14 种 multi-agent 失败模式。每个任务都需要一个 Harness：Claude Code 动态工作流深度解析https://cloudcold.ai/posts/2026-06-06-harness-claude-code/Sat, 06 Jun 2026 00:00:00 +0000https://cloudcold.ai/posts/2026-06-06-harness-claude-code/Anthropic 把 agent 编排计划从 context window 搬进 JavaScript 脚本，用代码确定性替代模型实时判断。深度解析运行时拓扑、三个强化机制、竞品横向对比、Multi-Agent 性能陷阱（-70%）、~15x token 成本、历史类比。Claude Code Dynamic Workflows: 从单兵到千军万马的编排革命https://cloudcold.ai/posts/2026-06-04-claude-code-dynamic-workflows/Thu, 04 Jun 2026 00:00:00 +0000https://cloudcold.ai/posts/2026-06-04-claude-code-dynamic-workflows/Anthropic Dynamic Workflows 全面解读：脚本驱动千军万马的编排架构，性能提升 90.2%，token 成本暴涨 15 倍。三层编排对比、竞品横向分析、历史类比、成本现实与关键风险。