Chinese contentWebflow、TELUS 等 4 支团队分享 AI Agent 生产化经验
Webflow、Bosch 这些公司在 n8n 的活动上讲怎么把 Agent 从 Demo 推到生产，四个真实团队踩坑实录，做落地的值得听。
#MCP and tools
1094 stories · 133 in 7 days · 463 in 30 days · Topic watch →
Oct 9
vLLM Semantic Router Adds Decision 2.0 Single-Pass Classification
Batching several classification questions into one forward pass cuts inference calls, which matters for anyone running routers or guardrails at scale on vLLM. Developers building model-selection or category-routing pipelines can adopt it directly in their existing Semantic Router configs. It competes with ad hoc multi-prompt setups that pay per-question latency.
Chinese content逆向工程 MCP 服务器 REA 获 29k stars，可让智能体分析任意应用
看这个 REA 项目，一个 MCP 服务器，没源码也能让 AI 智能体把 app 的机器码还原成可读代码讲清功能逻辑。
Stanford CS146S releases Week 3 on Agent Skills and CLI
The most actionable idea in the notes: mine your own chat logs for moments where you corrected an agent — each intervention point is a skill worth encoding as markdown. That turns personal friction into a portable, cross-project skill set, and the progressive-disclosure advice addresses why most skills fail (never triggered, or bloated into context overhead). Practitioners building agent workflows with Claude or similar coding agents can apply the SKILL.md patterns immediately.
REA v6.1.0 brings reverse engineering into agent MCP toolkits
The interesting shift is that reverse engineering, once a specialist skill, becomes a routine agent call with evidence-backed output rather than model guesses. Developers doing security audits, malware triage or app analysis can wire it into any MCP-compatible agent and keep everything on-device. Its coverage breadth, from APKs to EVM bytecode, compares favourably with narrower single-target tools.
Chinese contentAlixPartners 用 Claude Code 分析合同数据，防止成本节省重复计算
AlixPartners 两位顾问拿 Claude Code 跑合同和支出数据，讲怎么避免智能体把同一笔节省算两遍，做财务分析的很实用。
Chinese contentMercury Decide 登陆 OpenRouter，新增零数据保留付费端点
Inception 的 Mercury Decide 在 OpenRouter 上新增了零数据保留付费端点，输入 $0.02/M 打五折，输出还免费，适合对数据隐私有要求的场景。
Chinese contentEmad Mostaque 分享让 Grok 智能体与其他智能体互通的思路
Mostaque 分享了怎么让 Grok 智能体跟别的智能体对话，做智能体互操作的人可以看看思路。
Chinese contentGoogle 推出 Gemini Agent：每个任务配独立电脑，可选 Gemini 4 Argon 或 Claude Opus 5.5
Google 这个 Gemini Agent 给每个任务单独一台电脑，还能像员工一样挂邮箱进公司通讯录，模型还能在 Gemini 和 Claude 之间随便换，挺有意思。
Chinese contentAmazon Bedrock AgentCore 支付功能让智能体按次付费调用模型推理
Amazon 给智能体加了按次付费能力，Incarna 的智能体能用 x402 一单一单付钱给 BlockRun，几周集成活缩到几天。
Chinese contentZeta Global 发布 AthenaOS 与 AIM 推理模型，联合 Fireworks AI 打造
营销公司 Zeta Global 搞了个企业级智能系统 AthenaOS，还用 NVIDIA Nemotron 训了自家推理模型 AIM，底层靠 Fireworks AI，想看企业怎么自建模型可以瞧瞧。
Chinese contentdair-ai 推出 MCP 工具，让 Codex/Claude/Grok 机器人检索精选 AI 论文
dair-ai 把自己多年精选的 AI 论文库做成了 MCP 工具，接上 Claude 或 Codex 就能让机器人帮你查论文、写综述，做研究可以试试。
Oct 8
Chinese content斯坦福 CS 329Z 智能体工程课讲义公开，新增 Tool Use 等讲次
斯坦福的智能体工程课讲义全公开了，第 4、5 讲讲工具调用和框架编排，做 Agent 的可以免费白嫖看讲义。
Chinese contentGeoNatureAgent：面向地理空间任务的工具型智能体评测基准发布
一个专门测地理空间智能体的开源基准，103 个任务打九个模型，Claude Sonnet 4 第一但 DeepSeek V3.2 便宜 11 倍，做 Agent 评测的可以拿来就用。
Chinese contentHeyGen MCP 让 Grok @Bot 每天自动生成以你为主播的新闻节目
把 HeyGen MCP 接到 Grok @Bot 里，每天早上自动用你的虚拟形象播报 X 新闻，0.68.1 版还支持做 PPT 和发邮件。
Chinese contentOpenAI 推出 ChatGPT 插件扩展功能
OpenAI 给 ChatGPT 插件加了新玩法：能从侧边栏启动、@ 提及调用、还能带文件查看器，做插件开发的可以去看看演示视频。
Chinese contentGPT-6 Luna Decisions 上线 OpenRouter
OpenAI 的 Decisions API 现在在 OpenRouter 也能用了，输入 0.1 美元每百万 token，输出免费，做路由选模型的应用可以试试。
Chinese contentLangChain 更新 Deep Agents 的 skills 机制
LangChain 改了 Deep Agents 的 skills 功能，想给智能体注入领域知识的话可以看看他们怎么做。
Chinese contentdeepagents 支持加载 skill 时动态挂载工具
Harrison Chase 发的，deepagents 现在能按 skill 动态加载工具，用 OpenAI 或 Anthropic 模型还不会破坏 prompt cache，做智能体的可以看看。
Chinese content用一条提示词在 ChatGPT 里批量生成 HeyGen 个性化视频邀请
HeyGen 的 MCP 接进 ChatGPT 后，一条提示词就能给每个宾客生成专属视频邀请，做活动通知的朋友可以试试。
LangChain Managed Deep Agents v0.9 Adds Schedules SDK
Per-run config means one deployment can serve multiple teams or repos with different models and MCP servers, cutting infrastructure overhead for teams running many agent workloads. Developers building autonomous or long-running agents should evaluate the Schedules SDK for self-scheduling follow-ups without external cron jobs. It positions LangChain against managed agent platforms like OpenAI's Assistants and Anthropic's Agent SDK.
Chinese contentClaude Haiku 5.5 发布，平均运行成本比 4.5 低约 75%
Anthropic 把 Haiku 5.5 成本降到 4.5 的四分之一左右，还首次加了 effort 调节，Sonnet 5.5 缓存读取价格也砍半，跑大量 Agent 的可以算算账了。
Chinese content用 Lambda Durable Functions 自动执行 AWS DevOps Agent 修复方案
AWS 官方教程，教你把 DevOps Agent 的诊断结果通过 Lambda Durable Functions 和 Bedrock 变成一键可批准的修复动作。
Oct 7
Chinese contentTavily 赞助 LangChain Interrupt London 大会
Tavily 会出现在 Interrupt London 现场，做智能体联网搜索的可以顺便去展位聊聊，看看它和 LangChain 怎么配合。
Chinese contentAlibaba Cloud 与 Sui Foundation 合作，让 AI 代理用链上支付调用云服务
阿里云上了 Sui 链，AI 代理能自己付钱买云服务，新加坡 Sui Basecamp 上刚宣布的
SpaceXAI Cookbook adds 5 new Grok API sample apps
The cookbook doubles as an architecture reference: the pattern of using Grok for decisions and Imagine models for execution, plus cross-shot consistency via image editing instead of regeneration, applies to any agentic media pipeline. Developers building voice agents or content workflows can clone the repos on GitHub and adapt them directly. The itemised per-run costs also make budgeting for video generation pipelines more concrete.
html-plan plugin turns Claude Code plans into reviewable HTML pages
Plans as interactive HTML rather than walls of Markdown cut review time and make the feedback loop structured instead of ad-hoc, which could reshape how teams gate agent coding work. Developers running Claude Code on real projects can install the plugin and start routing plan approvals through numbered decision points. It comes from the Anthropic community plugin marketplace, so setup is a two-command install rather than custom tooling.
llm-openai-decisions 0.1a0 plugin for OpenAI Decisions API
Decision-style APIs price inputs only, which makes cheap bulk classification and moderation workflows far more predictable to budget than chat-style models. Developers already using Willison's LLM CLI can add structured yes/no, choice, and score queries with a single plugin install. The 2.4x price gap versus Jev gives teams with high-volume image or text gating tasks a concrete cost comparison to weigh.
OpenAI Decisions API Launch
The Decisions API represents a significant shift toward more efficient real-time model selection, potentially reducing latency in AI applications. Developers building applications requiring rapid decision-making should evaluate this API against alternatives like Perplexity's offering, as the performance difference could impact user experience and operational costs.
Chinese contentMistral AI 成为 Sourcegraph 新客户，用 Code Search 管理代码库
Mistral AI 用上了 Sourcegraph 的代码搜索和 MCP server，管理自家所有仓库，想了解这类工具怎么被模型公司用的话可以看看。
Chinese contentEvery 团队用 Claude Managed Agents 在 Slack 搭建全公司共用智能体
Every 团队把自己的工作流塞进一个 Claude 智能体，全公司通过 Slack 共用，还开放给了订阅者，做法挺值得参考。
OpenAI Decisions API enters public beta for all developers
Routing logic moves from hand-written prompt pipelines to a first-party API call, which shortens the glue code agent builders maintain. Developers building multi-model or tool-switching applications can test it now in public beta and benchmark the 10x speed claim against their own Responses API setups. It also signals OpenAI competing directly with self-built router frameworks like RouteLLM.
Chinese contentAgentMail 推出 AgentID，让 AI Agent 用验证邮箱登录应用
AgentMail 出了个 AgentID，让 AI Agent 拿验证邮箱登录应用，应用方还能按主人限流和管理，做 Agent 集成的可以看看。
Chinese contentSierra 推出 Personal Agent Protocol 连接企业与用户个人智能体
Sierra 发了个新协议，专门管企业怎么跟用户自己的个人 Agent 打交道，做客服或 Agent 开发的可以看看。
Chinese contentClaude 新增 Google 文档协作功能：贴链接即可并排编辑
Claude 现在能直接打开 Google 文档、表格和幻灯片，贴个链接就能和它并排一起改，付费用户都能用。
Chinese contentAgentMail 推出 AgentID：给 AI 智能体专属登录身份
如果你也在搭自己的智能体，肯定遇到过它借你的账号登录或被注册页拦下的问题，AgentID 就是让智能体有自己的邮箱身份去登录，还能追溯到主人，开发者接入还有 3 个月免费额度。
Chinese contentGumloop 推出 Agent Browsers：为智能体补上无 API 场景的浏览器能力
Gumloop 上线 Agent Browsers，没有 API 或 MCP 的网站也能让智能体直接用浏览器操作，做一次就存成配方每周自动跑，做个人智能体的可以试试。
Chinese contentGumloop 推出 Agent Browsers，让智能体操作没有 API 或 MCP 的网页
Gumloop 发了个 Agent Browsers，专门搞定那些没有 API 和 MCP 的网页操作，还带 1Password 集成和会话回放，做自动化工作流可以看看。
Chinese content用 Ampersand 把智能体接入客户 Salesforce，省掉认证和重试工作
如果你给客户做 Salesforce 集成，Ampersand 帮你搞定认证、token 刷新和重试这些脏活，只需配置字段映射，省不少事。
