{
  "report_date": "2026-05-22",
  "title": "AI 日报 2026-05-22",
  "summary": "本轮重点集中在 agent 工程栈：Google 把 Gemini 3.5、Antigravity、WebMCP 和托管 agent 放到同一条开发链路；Vercel AI Gateway 连续接入 Qwen 3.7 Max 与 Grok Build 0.1；GitHub Copilot 加强任务路由与语义 issue 检索；AWS SageMaker 用 OpenAI-compatible API 降低迁移成本；Anthropic 继续把 Claude Managed Agents 的执行环境和私有 MCP 访问边界拆开。GitHub Trending 继续显示 coding-agent skills、代码图谱、MCP 和 agent memory 工具升温；Builder feed 本轮受网络限制未收录。",
  "main_items": [
    {
      "title": "Google I/O 2026 把 Gemini 3.5、Antigravity 和 WebMCP 放进同一条 agent 开发链路",
      "event_date": "2026-05-19",
      "url": "https://developers.googleblog.com/en/all-the-news-from-the-google-io-2026-developer-keynote/",
      "source": "Google Developers Blog",
      "tier": "T0",
      "entities": [
        "Google",
        "Gemini 3.5",
        "Antigravity",
        "WebMCP",
        "Managed Agents"
      ],
      "summary": "Google I/O 开发者主题演讲确认 Gemini 3.5 系列、Antigravity 2.0、Antigravity CLI、Gemini API Managed Agents、WebMCP、Android CLI 和 Chrome DevTools for agents 等能力同场发布，核心不是单点模型，而是让 agent 在 IDE、浏览器、移动端和托管运行环境之间形成更完整的执行链路。",
      "bullets": [
        "Antigravity 2.0 与 Antigravity CLI 支持专用 subagents、跨平台终端沙箱、凭据 masking 和 Git policy hardening，说明 coding agent 的边界正在从模型选择扩展到执行环境和策略约束。",
        "Google AI Studio、Cloud Run、Firebase 与 Antigravity 的项目状态导出被放在同一段落，意味着从原型到部署的 agent workflow 会被更紧地绑定到 Google 开发者工具。",
        "WebMCP origin trial、Chrome DevTools for agents 和 Modern Web Guidance 指向浏览器侧 agent 工具标准化，后续值得跟踪它与现有 MCP server 生态的兼容边界。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    },
    {
      "title": "Vercel AI Gateway 接入 Qwen 3.7 Max 与 Grok Build 0.1，并把 Chat SDK 工具层开放给 agent",
      "event_date": "2026-05-21",
      "url": "https://vercel.com/changelog",
      "source": "Vercel Changelog",
      "tier": "T0",
      "entities": [
        "Vercel",
        "AI Gateway",
        "Qwen 3.7 Max",
        "Grok Build 0.1",
        "Chat SDK"
      ],
      "summary": "Vercel 在 48 小时内连续更新 AI Gateway 与 Chat SDK：Qwen 3.7 Max 可通过 `alibaba/qwen-3.7-max` 调用，Grok Build 0.1 可通过 `xai/grok-build-0.1` 调用，Chat SDK 新增 `chat/ai` 工具入口，`createChatTools(chat)` 可把聊天读写动作接入 agent。",
      "bullets": [
        "Qwen 3.7 Max 被 Vercel 定位为 agent foundation，覆盖 coding、office workflow automation 和 long-horizon autonomous execution。",
        "Grok Build 0.1 是 early access 的 agentic coding 模型，并且明确没有可配置 reasoning effort 或 non-reasoning mode。",
        "Chat SDK 工具层默认对写操作保留 approval gate，`reader`、`messenger`、`moderator` preset 也让 agent 权限范围更容易被产品代码表达。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    },
    {
      "title": "GitHub Copilot 同日更新任务路由、语义 issue 检索和 Web 模型列表",
      "event_date": "2026-05-20",
      "url": "https://github.blog/changelog/2026-05-20-auto-model-selection-now-routes-based-on-your-task-in-vs-code/",
      "source": "GitHub Changelog",
      "tier": "T0",
      "entities": [
        "GitHub Copilot",
        "VS Code",
        "semantic issue search",
        "model routing"
      ],
      "summary": "GitHub Copilot 的 auto model selection 现在会结合实时可用性、可靠性和任务维度为 VS Code 请求选模型；同日 Copilot Chat on web 上线语义 issue 检索，并收窄 Web Chat 的可用模型列表以保证稳定响应。",
      "bullets": [
        "自动路由会评估 reasoning、代码生成复杂度、bug 诊断难度和 tool orchestration 需求，管理员模型策略仍会生效。",
        "语义 issue 检索使 Copilot Chat 能用自然语言寻找、分组和分析 issue，不再只依赖标题或精确关键词。",
        "Web Chat 移除所有 Gemini 模型和部分其他模型，显示大厂 coding assistant 正在同时追求模型多样性和产品面稳定性。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    },
    {
      "title": "Amazon SageMaker AI endpoints 增加 OpenAI-compatible API",
      "event_date": "2026-05-20",
      "url": "https://aws.amazon.com/blogs/machine-learning/announcing-openai-compatible-api-support-for-amazon-sagemaker-ai-endpoints/",
      "source": "AWS Machine Learning Blog",
      "tier": "T0",
      "entities": [
        "Amazon SageMaker AI",
        "OpenAI SDK",
        "LangChain",
        "Strands Agents"
      ],
      "summary": "AWS 宣布 SageMaker AI real-time inference endpoints 支持 `/openai/v1` 路径，现有 OpenAI SDK、LangChain 或 Strands Agents 应用可以通过更换 endpoint URL 调用自有 SageMaker 模型，且支持 streaming 与 time-limited bearer token。",
      "bullets": [
        "该能力把自托管模型服务和 OpenAI-compatible client 连接起来，降低企业把 agent 或 LLM 应用迁移到自有 AWS 账户内推理端点的代码成本。",
        "SageMaker 通过 endpoint name 路由请求，并允许 inference components 下多个模型共享同一接口，适合多模型托管和 fine-tuned model serving。",
        "对平台团队来说，关键变化是不用在应用层维护 SigV4 wrapper 或自定义 client，就能把治理、部署和监控留在 SageMaker。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    },
    {
      "title": "Claude Managed Agents 增加 self-hosted sandboxes 和 MCP tunnels",
      "event_date": "2026-05-19",
      "url": "https://claude.com/blog/claude-managed-agents-updates",
      "source": "Claude Blog",
      "tier": "T0",
      "entities": [
        "Anthropic",
        "Claude Managed Agents",
        "MCP",
        "self-hosted sandboxes"
      ],
      "summary": "Anthropic 宣布 Claude Managed Agents 可在用户控制的 sandbox 中执行工具，并通过 MCP tunnels 连接私有 MCP servers；self-hosted sandboxes 处于 public beta，MCP tunnels 处于 research preview。",
      "bullets": [
        "官方说明 agent loop 仍在 Anthropic infrastructure 处理 orchestration、context management 和 error recovery，而工具执行移动到用户配置环境。",
        "self-hosting 控制代码执行位置，MCP tunnels 控制 Anthropic 如何访问企业网络中的 MCP servers，两者相互独立也可以组合使用。",
        "这会让企业评估 managed agent 时重点看凭据隔离、网络边界、审计日志和 sandbox provider 的实际运行特性。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    },
    {
      "title": "Claude Code 2.1.146 修复 Windows PowerShell、分页 MCP 和后台会话细节",
      "event_date": "2026-05-21",
      "url": "https://code.claude.com/docs/en/changelog",
      "source": "Claude Code Docs",
      "tier": "T0",
      "entities": [
        "Claude Code",
        "PowerShell",
        "MCP",
        "Agent SDK"
      ],
      "summary": "Claude Code 2.1.146 将 `/simplify` 更名为 `/code-review` 并支持 effort level，同时修复 winget 或 Microsoft Store 安装的 pwsh 命令行错误、MCP resources/prompts 分页丢项、Windows Terminal 后台会话 strobing 和 Agent SDK 流结束异常等问题。",
      "bullets": [
        "这次更新集中在工具可靠性和 Windows 体验，和本仓库最近遇到的 PowerShell / Windows 自动化边界高度相关。",
        "MCP 分页修复会影响大型 MCP server 的资源、模板和 prompt 枚举完整性，适合放入 agent 工具链稳定性观察。",
        "`/code-review high` 这种命令形态也表明代码评审开始被做成一等交互，而不是单纯的 prompt 模板。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    },
    {
      "title": "Hugging Face 发布 Ettin Reranker Family，补强小型 cross-encoder rerank 路径",
      "event_date": "2026-05-19",
      "url": "https://huggingface.co/blog/ettin-reranker",
      "source": "Hugging Face Blog",
      "tier": "T0",
      "entities": [
        "Hugging Face",
        "Sentence Transformers",
        "Ettin Reranker",
        "RAG"
      ],
      "summary": "Hugging Face 发布 6 个基于 Ettin ModernBERT encoders 的 Sentence Transformers CrossEncoder rerankers，并同时给出数据与训练 recipe，面向 RAG 检索链路中的轻量 rerank 评估和部署。",
      "bullets": [
        "发布覆盖 17M 到 1B 级别的 reranker artifact，给 RAG 团队提供了更细的延迟、成本和质量取舍空间。",
        "文章公开 distillation、training dataset 和 evaluation 细节，降低团队在自有语料上复现实验的门槛。",
        "相比泛化大模型更新，这类专用 reranker 更直接影响检索增强应用的召回后排序质量。"
      ],
      "image_urls": [],
      "importance": "major",
      "editorial_category": "ai_industry"
    }
  ],
  "model_releases": [
    {
      "name": "Gemini 3.5 Flash",
      "provider": "Google",
      "availability": "closed_api",
      "event_date": "2026-05-19",
      "url": "https://deepmind.google/models/model-cards/gemini-3-5-flash/",
      "source": "Google DeepMind Model Card",
      "summary": "Google DeepMind 发布 Gemini 3.5 Flash model card，并在 I/O 2026 开发者发布中把它作为 agentic coding 和 long-horizon agent 任务的重要模型。",
      "release_scope": "model_card_update",
      "image_urls": [],
      "importance": "notable"
    }
  ],
  "hot_blogs": [],
  "projects": [
    {
      "name": "anthropics/claude-plugins-official",
      "description": "Anthropic-managed Claude Code plugins 官方目录，在 GitHub Trending daily 与 Python daily 中出现，说明 Claude Code 插件生态正在被集中发现。",
      "url": "https://github.com/anthropics/claude-plugins-official",
      "event_date": "2026-05-22",
      "source": "GitHub Trending daily / Python daily",
      "signal": "trending",
      "domains": [],
      "importance": "notable"
    },
    {
      "name": "colbymchenry/codegraph",
      "description": "面向 Claude Code、Codex、Cursor 和 OpenCode 的本地预索引代码知识图谱，用于减少 token 和 tool call。",
      "url": "https://github.com/colbymchenry/codegraph",
      "event_date": "2026-05-22",
      "source": "GitHub Trending daily / weekly / TypeScript daily",
      "signal": "star_velocity",
      "domains": [],
      "importance": "general"
    },
    {
      "name": "ChromeDevTools/chrome-devtools-mcp",
      "description": "Chrome DevTools for coding agents，在 TypeScript daily 和全站 daily 中出现，适合关注 browser debugging 与 MCP 化工具链。",
      "url": "https://github.com/ChromeDevTools/chrome-devtools-mcp",
      "event_date": "2026-05-22",
      "source": "GitHub Trending daily / TypeScript daily",
      "signal": "trending",
      "domains": [],
      "importance": "notable"
    },
    {
      "name": "HKUDS/CLI-Anything",
      "description": "宣称让所有软件 agent-native 的 CLI-Hub 项目，在全站 daily 与 Python daily 中出现。",
      "url": "https://github.com/HKUDS/CLI-Anything",
      "event_date": "2026-05-22",
      "source": "GitHub Trending daily / Python daily",
      "signal": "trending",
      "domains": [],
      "importance": "notable"
    },
    {
      "name": "rohitg00/agentmemory",
      "description": "面向 AI coding agents 的 persistent memory 项目，在 GitHub Trending weekly 中出现。",
      "url": "https://github.com/rohitg00/agentmemory",
      "event_date": "2026-05-22",
      "source": "GitHub Trending weekly",
      "signal": "star_velocity",
      "domains": [],
      "importance": "general"
    },
    {
      "name": "Lum1104/Understand-Anything",
      "description": "面向 Claude Code、Codex、Cursor、Copilot 和 Gemini CLI 的代码图谱探索工具，在 TypeScript daily 中出现。",
      "url": "https://github.com/Lum1104/Understand-Anything",
      "event_date": "2026-05-22",
      "source": "GitHub Trending TypeScript daily",
      "signal": "trending",
      "domains": [],
      "importance": "notable"
    }
  ],
  "builder_observations": [],
  "schema_version": 1,
  "canonical_url": "https://jasonxzwen.github.io/ai-daily-cn/reports/2026/05/2026-05-22.html",
  "html_path": "reports/2026/05/2026-05-22.html",
  "generated_at": "2026-05-21T18:40:13.114Z",
  "hero_highlights": [],
  "github_trending": [],
  "stories": [
    {
      "story_id": "story-all-the-news-from-the-google-io-2026-developer-keynote",
      "title": "Google I/O 2026 把 Gemini 3.5、Antigravity 和 WebMCP 放进同一条 agent 开发链路",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-19",
      "primary_entity": "Google Developers Blog",
      "event_type": "launch",
      "object": "Google I/O 2026 把 Gemini 3.5、Antigravity 和 WebMCP 放进同一条 agent 开发链路",
      "what_happened": "Google I/O 开发者主题演讲确认 Gemini 3.5 系列、Antigravity 2.0、Antigravity CLI、Gemini API Managed Agents、WebMCP、Android CLI 和 Chrome DevTools for agents 等能力同场发布，核心不是单点模型，而是让 agent 在 IDE、浏览器、移动端和托管运行环境之间形成更完整的执行链路。",
      "why_it_matters": "Google AI Studio、Cloud Run、Firebase 与 Antigravity 的项目状态导出被放在同一段落，意味着从原型到部署的 agent workflow 会被更紧地绑定到 Google 开发者工具。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "Google Developers Blog",
          "url": "https://developers.googleblog.com/en/all-the-news-from-the-google-io-2026-developer-keynote/",
          "type": "source"
        }
      ]
    },
    {
      "story_id": "story-changelog",
      "title": "Vercel AI Gateway 接入 Qwen 3.7 Max 与 Grok Build 0.1，并把 Chat SDK 工具层开放给 agent",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-21",
      "primary_entity": "Vercel Changelog",
      "event_type": "update",
      "object": "Vercel AI Gateway 接入 Qwen 3.7 Max 与 Grok Build 0.1，并把 Chat SDK 工具层开放给 agent",
      "what_happened": "Vercel 在 48 小时内连续更新 AI Gateway 与 Chat SDK：Qwen 3.7 Max 可通过 `alibaba/qwen-3.7-max` 调用，Grok Build 0.1 可通过 `xai/grok-build-0.1` 调用，Chat SDK 新增 `chat/ai` 工具入口，`createChatTools(chat)` 可把聊天读写动作接入 agent。",
      "why_it_matters": "Grok Build 0.1 是 early access 的 agentic coding 模型，并且明确没有可配置 reasoning effort 或 non-reasoning mode。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "Vercel Changelog",
          "url": "https://vercel.com/changelog",
          "type": "source"
        }
      ]
    },
    {
      "story_id": "story-2026-05-20-auto-model-selection-now-routes-based-on-your-task-in-vs-code",
      "title": "GitHub Copilot 同日更新任务路由、语义 issue 检索和 Web 模型列表",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-20",
      "primary_entity": "GitHub Changelog",
      "event_type": "launch",
      "object": "GitHub Copilot 同日更新任务路由、语义 issue 检索和 Web 模型列表",
      "what_happened": "GitHub Copilot 的 auto model selection 现在会结合实时可用性、可靠性和任务维度为 VS Code 请求选模型；同日 Copilot Chat on web 上线语义 issue 检索，并收窄 Web Chat 的可用模型列表以保证稳定响应。",
      "why_it_matters": "语义 issue 检索使 Copilot Chat 能用自然语言寻找、分组和分析 issue，不再只依赖标题或精确关键词。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "GitHub Changelog",
          "url": "https://github.blog/changelog/2026-05-20-auto-model-selection-now-routes-based-on-your-task-in-vs-code/",
          "type": "source"
        }
      ]
    },
    {
      "story_id": "story-announcing-openai-compatible-api-support-for-amazon-sagemaker-ai-endpoints",
      "title": "Amazon SageMaker AI endpoints 增加 OpenAI-compatible API",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-20",
      "primary_entity": "AWS Machine Learning Blog",
      "event_type": "update",
      "object": "Amazon SageMaker AI endpoints 增加 OpenAI-compatible API",
      "what_happened": "AWS 宣布 SageMaker AI real-time inference endpoints 支持 `/openai/v1` 路径，现有 OpenAI SDK、LangChain 或 Strands Agents 应用可以通过更换 endpoint URL 调用自有 SageMaker 模型，且支持 streaming 与 time-limited bearer token。",
      "why_it_matters": "SageMaker 通过 endpoint name 路由请求，并允许 inference components 下多个模型共享同一接口，适合多模型托管和 fine-tuned model serving。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "AWS Machine Learning Blog",
          "url": "https://aws.amazon.com/blogs/machine-learning/announcing-openai-compatible-api-support-for-amazon-sagemaker-ai-endpoints/",
          "type": "source"
        }
      ]
    },
    {
      "story_id": "story-claude-managed-agents-updates",
      "title": "Claude Managed Agents 增加 self-hosted sandboxes 和 MCP tunnels",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-19",
      "primary_entity": "Claude Blog",
      "event_type": "research",
      "object": "Claude Managed Agents 增加 self-hosted sandboxes 和 MCP tunnels",
      "what_happened": "Anthropic 宣布 Claude Managed Agents 可在用户控制的 sandbox 中执行工具，并通过 MCP tunnels 连接私有 MCP servers；self-hosted sandboxes 处于 public beta，MCP tunnels 处于 research preview。",
      "why_it_matters": "self-hosting 控制代码执行位置，MCP tunnels 控制 Anthropic 如何访问企业网络中的 MCP servers，两者相互独立也可以组合使用。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "Claude Blog",
          "url": "https://claude.com/blog/claude-managed-agents-updates",
          "type": "source"
        }
      ]
    },
    {
      "story_id": "story-changelog-2",
      "title": "Claude Code 2.1.146 修复 Windows PowerShell、分页 MCP 和后台会话细节",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-21",
      "primary_entity": "Claude Code Docs",
      "event_type": "update",
      "object": "Claude Code 2.1.146 修复 Windows PowerShell、分页 MCP 和后台会话细节",
      "what_happened": "Claude Code 2.1.146 将 `/simplify` 更名为 `/code-review` 并支持 effort level，同时修复 winget 或 Microsoft Store 安装的 pwsh 命令行错误、MCP resources/prompts 分页丢项、Windows Terminal 后台会话 strobing 和 Agent SDK 流结束异常等问题。",
      "why_it_matters": "MCP 分页修复会影响大型 MCP server 的资源、模板和 prompt 枚举完整性，适合放入 agent 工具链稳定性观察。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "Claude Code Docs",
          "url": "https://code.claude.com/docs/en/changelog",
          "type": "source"
        }
      ]
    },
    {
      "story_id": "story-ettin-reranker",
      "title": "Hugging Face 发布 Ettin Reranker Family，补强小型 cross-encoder rerank 路径",
      "importance": "major",
      "trend": "AI industry",
      "event_date": "2026-05-19",
      "primary_entity": "Hugging Face Blog",
      "event_type": "launch",
      "object": "Hugging Face 发布 Ettin Reranker Family，补强小型 cross-encoder rerank 路径",
      "what_happened": "Hugging Face 发布 6 个基于 Ettin ModernBERT encoders 的 Sentence Transformers CrossEncoder rerankers，并同时给出数据与训练 recipe，面向 RAG 检索链路中的轻量 rerank 评估和部署。",
      "why_it_matters": "文章公开 distillation、training dataset 和 evaluation 细节，降低团队在自有语料上复现实验的门槛。",
      "evidence_level": "primary",
      "sources": [
        {
          "label": "Hugging Face Blog",
          "url": "https://huggingface.co/blog/ettin-reranker",
          "type": "source"
        }
      ]
    }
  ],
  "huggingface_trending": [],
  "chinese_media_dynamics": [],
  "daily_tracking": [],
  "official_org_updates": [],
  "evidence_assets": [],
  "quality_status": {
    "status": "degraded",
    "public_note": "Some discovery coverage is degraded; this report may be incomplete.",
    "affected_sections": [
      "github_trending",
      "builder_observations"
    ],
    "degraded_events": [
      {
        "section": "github_trending",
        "message": "github_trending coverage is degraded and should be disclosed in the public report.",
        "severity": "degraded"
      },
      {
        "section": "builder_observations",
        "message": "builder_observations coverage is degraded and should be disclosed in the public report.",
        "severity": "degraded"
      }
    ]
  }
}
