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.
248 articles · Page 1 of 25
Filter by tag
- All articles
- A2A
- ADK
- AGENTS.md
- AI
- AI Agents
- AI Strategy
- API
- AX
- AXI
- Agent Experience
- Agent Skills
- Agentic AI
- Agents
- Angular
- AngularJS
- Architecture
- Automation
- Bash
- Browser APIs
- C2PA
- CLI
- CSharp
- Cloudinary
- Community
- Conference
- Content Credentials
- Core Web Vitals
- DevFest
- DevRel
- Developer Tools
- Docker
- Embeddings
- Enablement
- Engineering
- FDE
- Firebase
- Frontend
- Function Calling
- Games
- Gemini
- Ghost
- Google ADK
- Google chrome
- GraphQL
- Harness Engineering
- Image Formats
- Image Optimisation
- Jamstack
- JavaScript
- LLM
- Large Language Models
- Learn To Code
- Local-first
- Loop Engineering
- MCP
- MCP Apps
- MEAN
- Machine Learning
- Memory
- MongoDB
- Multilingual
- MySQL
- NLP
- Next.js
- NoSQL
- Node.js
- Open Source
- PHP
- Performance
- Polymer
- Product Strategy
- Production AI
- Progressive Web App
- Progressive Web Apps
- Protocols
- Provenance
- RAG
- React
- Recruitment
- Review
- RxJS
- SQLite
- Security
- Serverless
- Service Worker
- Sponsored
- SynthID
- Taipei
- Thought Leadership
- Tool Calling
- Tool Use
- Tooling
- Transformers.js
- TypeSafe
- TypeScript
- VMware
- Veo
- Vercel AI SDK
- Verification
- Vue.js
- Web
- Web Assembly
- Web Development
- Web Performance
- WebGPU
- WebMCP
- this || that
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.