Writing

• How I check agent-built solvers

  What I check when coding agents turn papers into solvers, experiments and papers again. A snapshot of how I work in autumn 2026.

  Oct 2026
• Pull instead of push

  The bottleneck in a routing solver is usually one inner loop. Turning it around, so that work is gathered rather than handed out, made it about 18× faster on hard instances, and 274× on the best case.

  Sept 2026
• The 1993 answer was right until it wasn’t

  A paper from 1993 measured two ways of solving multi-commodity flow and picked a winner. Thirty-two years later the answer flipped, not because anyone found a mistake, but because the machines got big enough to change what the bottleneck was.

  Sept 2026