Research · 4 pieces

Writing

Experiments on AI coding agents, run on real repos and counted by scripts, not by opinion. Every figure comes from a saved run.

  1. Half-lifeThe rule that didn’t fadeOne vitest rule, four Claude tiers, 256 KB of padding, and no decay.vitest AGENTS.md · 4 Claude tiers · $106 at list prices
  2. Delegation taxThe orchestrator’s paraphrase is the bugTen coding tasks, solo against orchestrator, and three handoffs that broke the work.FeatureBench · pandas · 10 tasks · Opus 5.5, both arms
  3. Don’tWhat ‘never’ costsOne instruction rewritten forwards, 196 runs, and a result that failed to repeat.zod + remeda · 196 runs · $137 on Sonnet
  4. PieceworkWhat the typechecker can’t seeSeven breaks injected into WordPress/gutenberg, and what four tools found.WordPress/gutenberg · 102 packages · 7 injected breaks