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.

The newsletter

Ctrl+Alt+Delegate

The three-finger salute for work you should not be doing yourself.

  1. MCP Events

    An MCP server only answers when asked. MCP Events is a draft extension that lets it tell an agent something happened, by poll, push or webhook. Try all 3 here.

  2. Most people ask AI for answers. Your power users ask it about the work.

    Most people type into an AI assistant the way they type into a search box, and get search-box results. The questions power users ask instead, about the reader, the decision and what's missing, and why they're easy to teach.

  3. Juniors used to learn the hard way. Now someone has to teach them on purpose.

    Engineering judgement used to come bundled with the struggle of getting code to work. Coding agents made the struggle optional, so the judgement has to be handed down deliberately. What seniors can pass on, what juniors should practise, and the evidence for both.

  4. Your AI rollout raised the ceiling. Now raise the floor.

    A licence rollout makes your best people better and leaves everyone else where they were. 2 ways companies raise the floor, why most organisations can't use either, and a one-question test for whether yours has moved.

  5. What happens when AI gets its own computer and stops waiting for prompts

    Close your laptop and almost every AI tool you use stops dead. Always-on bots get their own cloud computer, their own memory and their own triggers. How they're designed, and what that changes about handing work over.

  6. Searching English text in Spanish, and why "balompié" goes skiing

    A Spanish search finds the right English sentence, and a formal Spanish word for football finds a ski trip. What an embedding model reads, what it gives back, how to tell a real match from a guess, and what changes with a second model running locally.

  7. EmbeddingGemma 2 in the browser: one search box for photos, sound, video and notes

    How EmbeddingGemma 2 puts text, images, audio and video into one vector space, and what it takes to run it entirely in a browser tab on WebGPU: token budgets, a 2,700-token wall, Matryoshka truncation, offline search, and why q4 doesn't load on WASM.

  8. How Code Mode turns MCP tools into a TypeScript API

    Code Mode gives the model a TypeScript API instead of a list of tools, then runs the program it writes in a sandbox. How the pattern works, built from scratch on MCP, QuickJS and Gemini, and where it gets fiddly.

  9. Most people are using AI back to front

    People hand AI the thinking and keep the searching, reading and table-filling for themselves. It should be the other way round. What to hand over, what to keep, and why knowing your subject matters more now.

  10. 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.