Writing
What I learn in production, written down.
Deep pieces you can act on: demos that run live in your browser, code checked against the packages you would actually install, and honest notes on what is not ready yet.
18 August 2026
An Overview Is Not an Instruction
Researchers measured what AGENTS.md files do to coding agents: success rates did not improve and cost rose over twenty percent. Named commands are obeyed hard, repository overviews are skipped, and most files are made of the second kind. What to write instead, and why the file should get shorter.
- AI in production
- Agents and protocols
17 August 2026
Neutral Is Not Ready: The Agentic AI Foundation's Five Projects
Google's A2A joined the Agentic AI Foundation on 17 August. Five competitors have now each donated a piece of the same agent stack. What that changes is who owns the projects, not whether they are ready, and only one of the five is worth betting a product on today.
- Agents and protocols
16 August 2026
The Agent Is Not the Domain Controller
Agent in the middle or at the edges? Wrong question. The agent works like a new hire, on whatever nobody has written a rule for yet. Every loop needs a merge, and good rules keep taking work off the agent's desk.
- Agents and protocols
- AI in production
11 August 2026
An Agent System Is Four Decisions
An agent system is four decisions: who may act, where the person stays in control, how the action is called, where the model runs. Each layer has its own tools, and all four answers fit on one page, written before any purchase order.
- Agents and protocols
- AI in production
4 August 2026
The Chat Box Is a Log: Agentic UI for the Web
When software answered, chat was the interface. When software acts, chat becomes a log, and the write has to land in the product, visibly and reversibly. Two protocol families now exist for exactly that.
- AI interfaces
- Agents and protocols
28 July 2026
The Page as a Tool Surface: WebMCP
An agent that clicks through your interface is guessing. WebMCP lets the page declare its actions as callable functions. The protocol is early; the discipline it demands is not.
- Web AI
- Agents and protocols
21 July 2026
Contracts, Not Prompts: Chrome's Built-in AI APIs
The browser did not ship a chatbot. It shipped task APIs with documented contracts: detect, translate, summarise. How to choose between the named APIs, the Prompt escape hatch and the cloud, and what the contract lets you promise.
- Web AI
On Medium · 2026
28 May 2026
Designing for Intelligence, Part 2: Twelve Patterns, Four Clusters
Six more interface patterns, grouped so a team can audit an AI product on a Monday morning.
17 March 2026
The Plausible-Correct Gap: Why AI-Generated Code Looks Right Until It Isn't
Four structural patterns that explain why AI output fails in production even when it looks finished.
On Medium · 2025
11 October 2025
Designing for Intelligence: Six Patterns for AI-Driven Interfaces
How interfaces have to change when the system is no longer a form that waits for a click.
On Medium · 2024
29 March 2024
Decoding Large Language Models
What these models actually do when they write code, and where that help stops being reliable.
