GitHub Copilot’s next edit suggestions just got faster, smarter, and more precise. 🧠
It’s powered by new data pipelines, reinforcement learning, and continuous model updates designed specifically for in-editor workflows.
Learn more about the model that predicts your next move. ⬇️github.blog/ai-and-ml/gith…6
✨🏖️ New Sandboxes for Deep Agents let you safely execute arbitrary code and bash commands in a remote environment.
They're useful when you need a clean, reproducible environment, want to run parallel or long-running tasks, or need to keep your local machine safe from untrusted code.
In this video you’ll learn how to:
1. Set up a remote sandbox
2. Make a tool call to execute code
3. Send a response back to your Deep Agentyoutube.com/watch?v=CejntU…S
Live now on OpenRouter!
INTELLECT-3 pushes the frontier forward by opening up how high quality models are trained. Weights, code, environments, and a detailed writeup are available to all.
Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack
Achieving state-of-the-art performance for its size across math, code and reasoning
Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more
Evaluating Deep Agents: Here's what we learned
Deep agents can't be evaluated like simple LLM tasks. After building and testing 4 production agents over the past few months, we learned that evaluating deep agents requires:
1. Bespoke test logic for each datapoint — each test case has its own success criteria
2. Single-step evals to catch regressions at individual decision points
3. Full turn evals to validate end-to-end agent behavior
4. Multi-turn evals that simulate realistic user interactions
5. Clean, reproducible test environments
👉 Read the full blog post on evaluation patterns for deep agents:blog.langchain.com/evaluating-dee…F
This is why Stripe works so aggressively to get all the startups in each YC batch to use it. Once you've picked a solution for payments, if it ain't broke you're not going to fix it. And then their revenues automatically grow at startup rates.
Many teams spend hours every week doing research – sales, marketing, recruiting, product. It’s necessary, but time consuming.
With LangSmith Agent Builder, you can create an agent to handle this research for you. It can:
- Call Tavily to search multiple web sites
- Synthesize the findings
- Send a report to your preferred channel, like Slack or email
- Work on a schedule or on-demand
Learn how to create a research agent with LangSmith and @tavilyai in just a few minutes 👇
youtu.be/2GeTGEVyPfo/?u…I
MCP、RAG、NLWeb 与 HTML 的对决：不同 Agent 网络交互接口的效能与效率对比
来自德国 Mannheim 大学的研究，探讨了一个核心问题：AI Agent 到底应该“怎么看”网页，才能不仅看得懂，而且干活快、花钱少？
论文：arxiv.org/pdf/2511.23281
核心背景：Agent 的“阅读障碍”
在目前的网络世界中，绝大多数网站是为人类设计的（漂亮的排版、图片、复杂的 HTML 结构）。当 AI Agent 试图通过“阅读”原始 HTML 代码来完成任务时，它往往会感到吃力：
· 干扰多： HTML 里充满了广告、样式代码，很难提取核心信息。
· 费脑子（Token）： 以前 Agent 需要消耗大量 Token 去理解网页结构，导致速度慢且昂贵。
四位参赛选手 MCP、RAG、NLWeb 与 HTML
· HTML - “硬啃生肉” - Agent 直接像浏览器一样读取网页源代码。这是最传统但也最笨重的方法。
· RAG - “划重点”先把网站内容抓取下来，建成索引。Agent 只阅读与问题最相关的片段，不看全篇。
· MCP - “走专用通道” - 网站提供一套标准的 API 接口，Agent 通过调用这些标准工具（如“搜索商品”、“获取价格”）来交互。
· NLWeb - “直接对话” - 这是一个新颖的概念。Agent 直接用自然语言问网站：“你们这儿有啥显卡？”，网站用 JSON 回答。
实验设计与“未来”模型
研究团队构建了一个名为 WebMall 的测试平台，包含四个模拟电商网站。他们使用了 2025 年底的顶级模型进行测试，包括 GPT-5、GPT-5-mini、GPT-4.1 和 Claude Sonnet 4。
测试的任务涵盖了从简单的“查价格”到复杂的“寻找替代商品”和“完成结账”。
关键发现与数据解读
实验结果非常具有冲击力，彻底暴露了传统 HTML 方式的短板：
A. 效率与准确率的碾压
· HTML 完败： 直接读 HTML 的 F1 分数仅为 0.67，是表现最差的。
· 新接口大胜： RAG、MCP 和 NLWeb 的平均 F1 分数都提升到了 0.75 - 0.77。
· GPT-5 的统治力： 当使用最强的 GPT-5 模型配合 RAG 接口时，F1 分数高达 0.87，任务完成率接近 80%。
B. 成本与速度的巨大差异
· 省钱（Token 用量）： HTML 方式平均每个任务要消耗 24 万 Token（因为网页代码太长）；而其他三种方式仅需 4.7 万 - 14 万 Token。
· 省时（运行时间）： HTML 方式平均耗时 291 秒（近 5 分钟）；而其他方式仅需 50 - 60 秒，速度提升了 5 倍。
C. 最佳性价比组合
论文指出了一个有趣的结论：虽然 GPT-5 + RAG 效果最好，但从性价比（成本 vs 效果）角度来看，GPT-5-mini 配合 RAG 是目前的最佳平衡点。
深度洞察与行业启示
这篇论文对未来的 AI 开发和网站建设指明了方向：
· HTML 已是过去式： 对于复杂的 Agent 任务，直接让 AI 解析 HTML 是极其低效且昂贵的。未来的 Agent 开发应尽量避免这种方式。
· “AI 友好型”网站将成趋势： 如果网站想要被 AI 助理（如购物助手、搜索助手）更好地索引和使用，就必须提供 MCP 或 NLWeb 这样的接口，或者优化内容以适应 RAG。
· 模型能力的依赖性： 实验发现，MCP 和 NLWeb 这种通过 API 或对话交互的方式，对模型本身的推理能力要求更高。GPT-5 在这方面表现出的优势比在简单任务中更明显。
总结
这篇论文告诉我们，为了迎接 AI Agent 时代的到来，我们需要重构互联网的“门面”。 未来的互联网不仅仅是给人看的（HTML），更是给 AI 看的（MCP/NLWeb）。对于开发者而言，放弃让 Agent “硬啃”网页，转而构建更结构化的数据交互通道，是提升效率的必经之路。
the paper compares 4 ways for LLM agents to use websites for shopping tasks, like using raw HTML pages, RAG, MCP, or NLWeb.
Finds that agents using plain HTML pages are less accurate and much slower than the other 3 options.
It studies LLM based assistants that search across online shops for products, compare options, and complete checkout.
These agents use 4 styles, HTML clicking, a retrieval augmented generation RAG search index, Model Context Protocol MCP APIs, and a standardized NLWeb natural language endpoint.
HTML means filling forms and following links, while the other 3 jump straight to structured product data.
The authors build simulated shops and give agents tasks like product search, vague search, finding cheapest offers, and checkout.
They track how many correct products each agent returns, how often the answer is fully right, tokens spent, and runtime.
On average, RAG, MCP, and NLWeb reach about 0.75 F1, while HTML stays near 0.67 and misses more.
HTML agents also burn more compute, needing about 3x more tokens and 5x more time per task than others.
So the study recommends RAG agents, often with GPT5 or cheaper GPT5 mini, and keeping HTML browsing only as backup.
---
Paper Link – arxiv. org/abs/2511.23281
Paper Title: "MCP vs RAG vs NLWeb vs HTML: A Comparison of the Effectiveness and Efficiency of Different Agent Interfaces to the Web (Technical Report)"
Dify v1.10.1 is live
Major backend upgrade
- Dify now ships with full MySQL support, alongside PostgreSQL and OceanBase. This release adds multi‑database adaptation, refactored SQL and migration logic, and new configuration options to make the metadata layer more flexible, compatible and maintainable.
(with special thanks to the @OceanBaseDB team for their support)
- Performance: a much smoother Workflow Editor and runtime, even ~200‑node workflows remain responsive.
- Stability: many fixes across workflow execution, knowledge base 500s, Node.js SDK / OpenAPI loading, plugins, logging, webhooks, i18n and more, plus expanded tests and dependency upgrades.
Important for self‑hosters: Docker images now run as a non‑root user by default. If you use Docker Compose, please chown 1001:1001 ./volumes/app/storage before upgrading.
Full changelog:
github.com/langgenius/dif…
Learn how to build a deep agent using 3 model with @LangChainAI's open-source deepagents harness:
- Configure the "gemini-3-pro-preview" model using the `ChatGoogleGenerativeAI` import.
- Delegates research tasks to specialized sub-agents with isolated contexts.
- Utilizes native file system tools to manage tasks and write files.
- Autonomously plans workflows and saves comprehensive Markdown reports to files.
I'm lucky to work with @DavidGeorge83 at a16z! I learn so much from talking with him, and very happy to see this podcast finally out.
.@DavidGeorge83 leads @a16z Growth, which has backed some of the most successful technology companies including Stripe, SpaceX, Waymo, Databricks, Figma, OpenAI, and Cursor.
This conversation is a detailed look at how David built and runs the growth business. He shares how they win the most competitive deals, the principles behind its culture, and the framework he uses to evaluate companies.
We discuss how a16z is investing across the AI stack -- from model providers to applications -- and why this platform shift creates a real window for startups to displace incumbents.
Throughout the episode, David shares the models that guide his investing: why markets misprice consistent growth, what makes "pull" businesses so powerful, and why he likes to back a certain kind of founder he calls the "technical terminator."
It’s fun to talk with students of the game, and David is certainly one of them. He's spent his career studying great companies, markets, and founders, and you can feel that depth in how clearly he connects patterns across cycles.
Enjoy!
Timestamps:
0:00 Intro
6:54 Enterprise AI
10:23 Lessons from Waymo
16:17 Technical Terminators
22:34 The "Glengarry Glen Ross" Rule
31:23 Winning Competitive Deals
42:24 The "Yankees" Mindset
46:13 Decision Making
51:12 Model Busters
53:42 Push vs. Pull Markets
1:10:23 How Startups Beat Incumbents
1:13:38 The Kindest Thing
