What I’m focused on now.

Leading engineering at Roadie. Most of my attention goes to making the catalog useful to agents as well as people: openroadie, a live and permissioned graph of services, teams and dependencies, is the centre of that.

Viberglass, an open-source workspace where a team and its coding agents work on the same tasks. The current focus is evaluation: grading agent runs against real ticket outcomes, measuring reliability over repeated runs rather than single passes, and treating cost and latency as results, not footnotes.

Deliberately deepening in agent evaluation, harness and context design, inference economics, and the sandboxes agents run in. Each topic ends in something I can measure and write about.

Qwen3.8-Flash-Next, a 125B-parameter model, on my NAS at 40–48 tokens per second, served by Strata next to my file shares.

Getting back to it after a long pause. This site was rebuilt in October 2026, and the first new articles are about running large open models on hardware you already own, and how to evaluate coding agents without fooling yourself.