// blog
Writing about the things we build and break.
Nobody Marks Their Own Homework
At half past three in the morning a coding agent improved the harness it works in, sent its own change to an independent reviewer, reproduced the reviewer's finding and fixed it before the next scheduled run. No human wrote, reviewed or corrected any of it. This is how the harness is built.
A Job Should Survive Its Agent
A directory check stopped us reopening our own task.
The Cloud Was a Phase
What a local inference box suggests about the falling cost of capable models
The Workhorse Comes Home
A local inference box and the 28-second test that found a silent harness failure
And Then There Were Graphs
We Took a Twelve-Word Meme Seriously. It Cost Five Million Tokens.
Verified Public Facts
Beyond Model Leaderboards
A Protocol Paper and Design History for Auditable Coding-Agent Tournaments
Beyond Model Leaderboards: The Operational Supplement
Adaptation Checklist, Reference Parameters and Publication Source Map
The King Is Dead. Long Live the King. Until Tuesday.
What a 608-Row Tournament, a 28-Assignment Shakeout and 96 Follow-Up Executions Taught Us About Building an AI Workshop at Frontier Speed
When Washington Is Your Landlord
How to Survive the Geopolitical Enclosure of Frontier Models