Can you make AI smarter by upgrading the tools and instructions around it, without touching the model itself?

This explores whether improving the scaffolding around a frozen model (prompts, tools, context handling, control flow, together called the 'harness') is a separate route to better AI than building a bigger or better-trained model. The corpus suggests it is separate in mechanism but not independent of the model in practice.

The gain comes from somewhere different. Can execution harnesses lift model performance without retuning weights? shows execution-system tuning lifting Terminal-Bench 2.1 accuracy across several models with no weight changes. The same runbook carried over to a newer model unchanged, reaching 95.3% on GPT-5.6 and adding 5.4 points to DeepSeek-V4 Flash. Can frozen models improve by evolving their harnesses? gets an average 17-point gain by letting harness variants compete while the model stays frozen. Model-side scaling, by contrast, follows smooth laws. How should finetuning scale with model and data size? finds finetuning scales multiplicatively, and larger base models help more than extra data. Harness gains come from search and selection over code and instructions rather than gradient descent, and the corpus has no comparable curve for them. 'Scaling' here means running more optimization, not following a law.

It also buys something different. The biggest lever isn't making the model think harder. It's taking fragile reasoning out of the model. Can a stronger model lift a weaker one at test time without retraining? nearly doubled a weaker model's Theory-of-Mind performance by moving unstable reasoning into deterministic code and task-specific routing, not by encouraging longer reasoning. Can explicit behavior maps help weaker planners compete with stronger models? let weaker planners match stronger models at locating the right code, just by handing them a map from behavior to code. Can agent harnesses be automatically optimized across many environments? found four mechanisms that cut token traffic by 44.7–49% at comparable performance, which is a cost win rather than a capability win. It also suggests harness gains are orthogonal to model improvements.

Model scaling already had an escape hatch from 'just make it bigger'. Can inference compute replace scaling up model size? showed inference compute can trade off against parameters on hard prompts, so model size was never the only dial. Harnesses look like another dial in that family. They can also be improved in independent pieces: Can harness modules improve separately from benchmark data? evolves harness modules separately, and the gains held on unseen tasks and domains.

The two paths are still coupled. Do stronger models always evolve harnesses better? found that models of every tier write equally useful harness edits, but the ability to benefit from those edits follows an inverted U. Weak models fail to invoke the harness, and strong models struggle to follow its instructions faithfully, so mid-tier models gain the most. Harnesses also don't port freely: Should safety harnesses be customized for each deployment? shows a harness strict enough for one model over-blocks another. Where does model adaptation actually happen? gives the cleanest framing: improvement belongs to the loop linking model, harness and external contract, not to either piece alone.

So harness scaling is a different path, and a cheaper one to iterate on. It works on frozen weights, and new models can inherit its gains. But it complements model scaling rather than replacing it. The corpus doesn't say where harness gains plateau, or whether they shrink as base models get stronger. The inverted U hints they might.