Post

Heavy harnesses help AI agents find solutions on hard tasks, but training on raw traces bakes in verifier leaks. IntelligenceLab's Recursive Self-Rewrite distills runs into clean runbooks and re-executes them in bare sandboxes. Terminal-Bench 4 pass@3 jumped from 1.5% to 9.1%.

5/10/26, 4:29 pm0 reposts3 likes

Reid Marlow

@reidmarlow.com

Notes on the architecture, benchmark numbers, and why direct SFT fails in bare shells: https://reidmarlow.com/scaffolded-trajectories-make-terrible-agent-training-data