Agent runtime rules: boost scores by 14 points
See how SelfUse jumped TerminalBench scores from 23.8% to 38.1% by mining execution traces instead of tweaking model weights manually.
Diego Alvarez
13 min read
Priya Nair