Meta-Harness: Chelsea Finn et al. Show Automated Harness Optimization Beats Hand-Engineering on IMO Math and Coding

arXiv / Matthew Berman·high signal

arXiv 2603.28052 from Yoonho Lee, Chelsea Finn, and Omar Khattab introduces Meta-Harness — an agentic outer loop (using Claude Code) that searches over harness code by reading source, scores, and execution traces of prior candidates via filesystem. On text classification it beats SOTA context management by 7.7 points while using 4x fewer tokens; on 200 IMO-level math problems it gains 4.7 points across five held-out models; on TerminalBench-2 agentic coding it surpasses all hand-engineered baselines. The implication for builders: once the search space is codified, general-purpose agents outperform manual harness tuning.