Building systems
that last

We pre-registered a self-improvement experiment, and it was over in one generation — the model aced every task and the selection gate became mathematically unsatisfiable. Two weeks later another lab hit the same wall. The instrument didn't fail. The instrument class did.

July 21, 2026 · 5 min read

In six months the field converged on a belief — the harness, not the model, is what matters — and backed it with striking empirics. What it doesn't have is a single theorem about when the loop it now worships converges, oscillates, or collapses. A map of the gap, and the four results that would fill it.

July 21, 2026 · 6 min read

We took Houston — an MIT-licensed agent runtime — and grafted a new brain onto it. In a single hack camp the fork went 89 commits ahead of upstream: four net-new Rust crates, a full Linear integration, a feature-flag platform, a push pipeline. The features are the easy part to show. The real story is the method that produced them — autonomous agents working in parallel git worktrees under a governance harness that made the speed safe instead of reckless.