On the bench
on the bench
- factoryrunning nightlyA Linear task goes in, a pull request comes out. Local models, deterministic gates. phile 06
- quality-labbuildingEvals made from my own merged pull requests, to choose a harness with numbers.
- hiveloomwith a friendA harness that keeps small models on track. get.hiveloom.cloud
- open sourcemergedSmall fixes to inspect_ai, inspect_evals, MTPLX and lince, each with a test. github
hardware
- dgx spark × 2The factory runs here: one model at a time, served with vLLM across both.
- macbook pro m5 max128GB of unified memory, for the local experiments.
- thinkpad e14The everyday machine.
metrics
- tokens used, since 2023307B
- prompts I typed14,000+
- with "ok"376
- with "continue", "continua" or "procedi"273
- with "merge"194
- with "grazie"51
- agents I work with6
- claude code, codex, opencode, pi, hermes, hiveloom
- agents running at the same time, peak107
- subagents spawned1,300+
- longest session without a break15.9h
- busiest day128 prompts
Updated 2026-10-20.