Evolved harnesses transfer across models — a single optimized harness improves five different LLMs

A single evolved retrieval harness, optimized on one model for IMO-level math problems, improved performance by an average of 4.7 percentage points (34.1% to 38.8%) across five completely different held-out models. The harness was never trained on these models — it transferred directly.

This has a profound implication for the investment calculus of Agents learn at three distinct layers — model weights, harness code, and context configuration: if harness improvements transfer across models, then the harness layer has even higher leverage than the model layer — you invest once and benefit regardless of which model you swap in. It strengthens A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one with hard evidence: the harness isn’t just important for one model, it’s a model-agnostic amplifier. It also explains why Build for the model six months from now, not the model of today isn’t as scary as it sounds — a well-evolved harness survives model upgrades, while model-specific scaffolding (the Scaffolding is tech debt against the next model — the bitter lesson applied to product building problem) does not. The distinction matters: harness improvements that encode domain patterns and evaluation structures persist; scaffolding that patches model weaknesses expires.