Do the scaffolding, tools and prompts wrapped around an AI help average-strength models more than the very best ones?

This asks whether the scaffolding around a model helps mid-tier models more than frontier ones. The corpus's most direct evidence says yes, with a twist. One study found that a model's ability to write good harness updates is flat across capability tiers. Its ability to benefit from those updates follows an inverted U that peaks in the middle: Do stronger models always evolve harnesses better?. Weak models often fail to invoke the harness at all. Strong models struggle with faithful instruction-following, so they drift from the prescribed steps. Being smarter doesn't mean using the scaffolding better.

The weak end isn't hopeless, though. It depends on how the harness is designed. A behavior-to-code map let weaker planners match stronger models at finding the relevant code, while using 8-13% fewer tokens: Can explicit behavior maps help weaker planners compete with stronger models?. In another case a stronger model built an inference-time harness that nearly doubled a weaker model's Theory-of-Mind score. It did this mostly by moving shaky reasoning into deterministic code and routing tasks, not by asking for longer thinking: Can a stronger model lift a weaker one at test time without retraining?. Both harnesses take decisions away from the model instead of asking it to make good ones. That suggests, as my reading and not a claim from the notes, that where the peak falls depends on how much judgment the harness demands of the model.

Frontier models still gain. A runbook tuned around frozen weights reached 95.3% on GPT-5.6 on Terminal-Bench 2.1 and lifted DeepSeek-V4 Flash by 5.4 points: Can execution harnesses lift model performance without retuning weights?. Evolving harness variants with the model frozen gave an average 17-point gain across benchmarks: Can frozen models improve by evolving their harnesses?. Neither result is split by model tier, so they can't say who gains most. A model already near 95% also has little room left to gain, and the notes don't test that directly.

Two cautions apply. First, harness gains have to be measured against a baseline that gets the same search and inference budget. Otherwise you may be crediting the harness for what extra computation would have bought anyway: How should we measure gains from automatic harness evolution?. Second, harness improvements seem to be a separate lever from model improvements. Automated optimization across many environments found mechanisms that cut token traffic by about half at comparable performance: Can agent harnesses be automatically optimized across many environments?. Routing queries among specialized models is another lever that can beat a single frontier model: Can routing beat building one better model?.

The corpus suggests mid-tier models are the sweet spot for harness gains. They are capable enough to follow the scaffolding but not so strong that they stop following it faithfully. The corpus has one direct study of this and no head-to-head comparison across tiers, so treat it as a strong hint and not a settled result.