Does harness benefit depend on which model tier you use?

This explores whether harness benefit is uniform or tier-dependent, and the corpus says it is sharply tier-dependent — but in a counterintuitive shape. The cleanest finding is that the ability to *write* useful harness updates is roughly flat across model tiers, while the ability to actually *benefit* from those updates follows an inverted U: it peaks at mid-tier models Do stronger models always evolve harnesses better?. Weak models fail because they don't reliably invoke the scaffolding in the first place; frontier models underperform expectations because they struggle to follow the harness's instructions faithfully, preferring their own judgment. So the model that gains most from a harness isn't the strongest one — it's the one capable enough to use scaffolding but not so capable it overrides it.

That reframes the broader claim that agent reliability migrates from model weights into the surrounding harness Where does agent capability really come from?. The migration isn't a free lunch you can pour onto any model — there's a capability window where externalized memory and protocols pay off most. This rhymes with how prompt techniques behave across tiers: rephrasing and background-knowledge prompts lift cheap models, while step-by-step reasoning prompts actually *reduce* accuracy in high-performance models, which already do that reasoning internally Do prompt techniques work the same across all LLM tiers?. In both cases, scaffolding substitutes for a capability the model lacks — and stops helping (or starts hurting) once the model already has it.

There's a second, more architectural sense in which the answer is yes: harness design can deliberately assign different jobs to different tiers rather than asking one tier to do everything. Hierarchical RAG routes filtering, pruning, and citation to cheap models like Gemini Flash while reserving an expensive model for final synthesis — and beats uniform deployment on both cost and quality Can smaller models handle RAG filtering while larger models focus on synthesis?. The benefit here isn't 'does the harness help this model' but 'which tier should sit in which slot of the harness.' Weak models can even match strong ones inside the right scaffold — but only when the harness supplies a verifiable soundness signal (tests, proofs, type checks) that converts their many guesses into correct selections When can weak models match strong model performance?.

Underneath all of this is a deeper rule the corpus keeps surfacing: what a model can gain from any external intervention is bounded by what's already latent in it. Spurious or random rewards boost reasoning dramatically in Qwen2.5-Math but do nothing for Llama and OLMo, because the gain depends on pretraining having planted the relevant behavior Why do random rewards improve reasoning for some models but not others?. Similarly, which specialization techniques are even available is capped by your access tier — black-box, grey-box, or white-box Does model access level determine which specialization techniques work?. The harness, the prompt, and the reward are all amplifiers, not sources.

The thing you didn't know you wanted to know: the model most worth building a rich harness for is the mid-tier one, not the flagship. The frontier model can author the scaffolding for everyone else, but it's the capable-yet-compliant middle that turns that scaffolding into the biggest reliability gain — which is a strong argument for cost-efficient agent fleets over dumping everything on the most expensive model.