Sébastien Chopin

Most knowledge agents start the same way. You pick a vector database, then build a chunking pipeline. You choose an embedding model and then tune the retrieval parameters. I recently authored this piece on the @Vercel blog, breaking down a different and better way you can create knowledge agents. Powered by Vercel Sandbox, Workflows, Chat SDK, and more. h/t @hugorcd for creating the beautiful template that led to the creation of this blog post. vercel.com/blog/build-kno…

Just shipped something I'm really excited about in @𝗇𝗎𝗑𝗍𝗃𝗌/𝗆𝖼𝗉-𝗍𝗈𝗈𝗅𝗄𝗂𝗍: Code Mode. The idea: instead of the LLM calling your MCP tools one at a time (and resending ALL tool descriptions every single round-trip), it writes JavaScript that orchestrates everything in one go. Loops, conditionals, Promise.all, real control flow, not 8 separate LLM turns. With 50 tools the token savings are insane: -81% on tool description overhead alone. And the best part? Your existing tools don't change at all. One line: `experimental_codeMode: true` Runs in a secure V8 sandbox thanks to @rivet_dev's secure-exec, perfect timing with their launch yesterday. mcp-toolkit.nuxt.dev/advanced/code-…

Every month I periodically see the recycled take that “something better than chat” is coming for AI. That chat is temporary. In fact, I predict the opposite. More of our work and life will happen through chat and voice interfaces of increasingly intelligent agents. 🦞 OpenClaw is the lobster in the coal mine of this. It’s true that as humans we consume with all the senses. When we use our internal company agent at @vercel, it can answer any question in English, but also plot data and richly visualize it. All in the chat medium. You start with any question, and receive periodic reports, much like an enterprise claw, on @slackhq. If the chat is not enough, you’re one click away from more depth and refinement on a web page. I don’t believe web pages are going anywhere. Many will evolve to accept natural language, which is the lingua franca of AI, and stream both text and complex data. We call this Generative UI (eg: json-render.dev). And web pages will also be crucial in enriching the conversations that happen elsewhere, whether you’re chatting with your agent on WhatsApp or Slack. We’re building chat-sdk.dev to make this universal interface as easy to deploy as possible.

.@nodejs has always been about I/O. Streams, buffers, sockets, files. But there's a gap that has bugged me for years: you can't virtualize the filesystem. You can't import a module that only exists in memory. You can't bundle assets into a Single Executable without patching half the standard library. That changes now 👇