Skip to main content

Writing

Notes on AI, agents, web technologies, and building with LLMs

Technical articles on building with LLMs: agents and agent harnesses, MCP, evaluation, and the web technologies underneath. Worked examples with code, written for working developers.

  1. Jev, the model that can't write

    Jev is a model that cannot generate a single word of text. It picks from options you give it and returns a probability for each, dozens of questions per call, in a few hundred milliseconds. To see what that's good for I gave it a gamebook to referee: you type anything, Jev decides what you meant, whether it would work, and what it says about you.

  2. MCP, WebMCP, A2A and ARD

    Four protocols cover almost everything an agent needs to reach: a tool, a web page, another agent, and a catalogue of what exists. MCP and WebMCP are separate specifications by different authors. ARD is four months old and indexes the other three. Here's what each one is and who's on each end of the wire.

  3. Musubi: a history of everything you asked

    I have thousands of conversations with AI tools scattered across four applications. Musubi reads all of them, distils them, and hands the result back to any agent over MCP, with nothing ever leaving the machine.

  4. Stop asking the model to remember arithmetic

    Most of what is in my skill files should not be there. A skill is knowledge the model needs while it works; a check is a fact about whether the output came out right. Almost everything I had written into skill files was the second kind wearing the clothes of the first.

  5. When to write an agent skill

    Vercel ran an eval where installed skills scored exactly the same as no documentation at all, because the agent never opened them in 56% of cases. A flat docs index in AGENTS.md scored 100% on the same suite. The useful reading of that result is about which of two jobs you're doing: putting a fact in front of the model, or getting a sequence carried out the same way every time.

  6. Routing a prompt between Gemini Nano and the cloud

    Chrome and Edge both ship a language model, and the AI SDK makes swapping one for a cloud model a one-line change. The interesting part is the router that decides which one runs, and every constant in it comes from a measurement.

  7. MCP's Progressive Tool Discovery

    Progressive tool discovery is usually sold on prompt caching. Keep the tools array fixed, the argument goes, and the cached prefix survives. I ran 72 benchmarks across gemini-3.5, 3.6 and 3.7 Flash at 14, 42, 84 and 280 tools. The fixed prefix cached in exactly zero of 36 runs, the pattern it argues against cached up to 93%, and the point where the pattern starts paying moved down to about 20 tools on the newest model.

  8. Stateless MCP servers

    The MCP server most tutorials ship, mine included, opens with a session handshake: initialize, get back an Mcp-Session-Id, carry it on every request, hold a live transport per client on the server. A stateless server drops all of it. Each request stands alone, any copy of the server can answer it, and the July 2026 spec makes that the only model there is. Here are both servers side by side, compared on the wire.

  9. The agent said it was in stock

    Agents keep giving me confident, wrong answers about the real world. WebMCP fixes half of that problem; discovery is still wide open.

  10. The forward deployed engineer land grab

    Four companies committed roughly $9 billion in nine weeks to putting their own engineers inside customer organisations. What a forward deployed engineer actually costs, who really gets one, and the question to ask before you sign anything: what remains when they leave?