Engineering
How the work actually gets built.
These are build-logs, not announcements. Each one traces how a product was built, the decisions behind it, the bugs that got past testing, and what was learned in the recovery. They are written in the same plain register as the software, and they lead with the mistakes, because the recovery is where the judgment shows.
13 June 2026 · ~16 min read
Orchestrating a fleet of AI agents
A working software engineer’s retrospective on coordinating sixteen AI coding agents across multiple products, the disciplines that survived contact with reality, and why the work amounts to what Addy Osmani recently named loop engineering, practiced here at fleet scale. Includes an honest on-ramp for anyone considering this kind of setup without sixteen terminals on the desk.
28 May 2026 · ~12 min read
Building ClaudeLink
An engineering retrospective on the open-source MCP server that lets multiple AI coding agents work as a team. The npm bin-resolution fight, a dependency default that flipped under me, the moment the platform safety model showed my autonomous-reply design was wrong, the bugs only real multi-model use could find, and the recovery watcher that took three tries to get right.
13 June 2026 · ~11 min read
The MDX brace trap
Ten ways an LLM-authored Markdown file silently breaks an MDX build, found over four months of shipping a content-heavy curriculum site. The classes range from set notation in prose to LaTeX subscripts, YAML colon-space, and a tenth class that does not exist in human-authored content. Plus the three-layer defense that finally made the build a gate instead of a guess.