7 min read
Writing your own coding harness
The loop at the centre of a coding agent is about fifty lines. Where the rest of the work went when I built a harness from scratch.
7 min read
How the KV cache shapes AI agent design
Prompt caching only works when each request begins with the same tokens as the last. The design rules that follow for AI agents, and how to check them.
5 min read
The economics of an agent's context window
An agent re-sends its whole conversation every turn. What that does to a bill, what a real project's numbers showed, and the five habits that changed it.
8 min read
Context engineering in practice
AI coding tools are context engines. How the harness works, the practices that improve what it produces, and what to settle before a team adopts them.