Does tidying how an AI's workspace is organized save more effort than upgrading to a bigger, pricier model?

This explores whether reorganizing the scaffolding around a model (the harness: how code, context and tool output are laid out for it) saves more planning tokens than moving to a bigger model. The corpus suggests structure is the cheaper lever, and for some jobs it can stand in for size. It does not work equally well for every model. The corpus has no head-to-head sweep of harness design against model size on token cost, so what follows is assembled from adjacent results.

The closest evidence is a harness repository reorganized around runtime behavior. It uses an explicit map from behaviors to the code that implements them. That cut planner tokens by 8–13% and raised win rates by 10–19 points. Weaker planners using the map matched stronger models' code localization on every precision and recall metric (Can explicit behavior maps help weaker planners compete with stronger models?). Without the map, a planner spends tokens, and intelligence, working out where things live. The structure pays that cost once, so a smaller model no longer needs to be bigger to find its way around.

The effect gets larger at the system level. Auto-research loops run across many environments found four mechanisms: action execution, context compaction, observation handling and delegated reading. Together they cut token traffic by 44.7–49.0% on a 51-task benchmark with comparable performance (Can agent harnesses be automatically optimized across many environments?). That figure is total traffic, not planner tokens alone, so it can't be compared directly with the 8–13% above, but both point the same way. A separate result shows harness optimization lifting several models on Terminal-Bench 2.1 with frozen weights. The same runbook carried over to newer models without changes (Can execution harnesses lift model performance without retuning weights?). Harness gains appear to stack on top of model gains rather than compete with them.

The catch is that a good harness does not rescue every small model. Skill at writing harness updates is flat across model tiers. Skill at benefiting from them follows an inverted U that peaks in mid-tier models. Weak models often fail to invoke the harness at all, and the strongest models struggle to follow its instructions faithfully (Do stronger models always evolve harnesses better?). So a weak planner behind a well-organized harness has a floor. One way around it is to spend model size once, at build time. A stronger model built inference-time harnesses that nearly doubled a weaker model's Theory-of-Mind performance. It did this mainly by moving unstable reasoning into deterministic code and task-specific routing, not by encouraging longer reasoning (Can a stronger model lift a weaker one at test time without retraining?). That result measures performance, not tokens.

Two results explain why tokens are a good thing to trim. More reasoning tokens don't just hit diminishing returns. They can make a model abandon answers it had already gotten right (Why does more reasoning sometimes make models worse?). And in a 115-day persistent-agent case study, 82.9% of tokens were cache reads, so the useful cost unit became completed artifacts rather than raw tokens (Do persistent agents really cost less per token?). The same principle shows up outside agent harnesses. Retrieving 8 of 32 user-model fields matched full-model fidelity at 211 tokens instead of 915 (Can a smaller user model subset match full model performance?). Deciding what the model sees often saves more than choosing how large the model is.