When you improve the code wrapped around an AI model, does the gain come from more compute, or from better feedback?

This explores whether harness scaling is driven by raw token spend or by compute spent closing feedback loops, and what makes that feedback effective. The corpus has no note that measures "feedback compute" directly, so the picture below is assembled from several. It points one way: most harness gains come from loops that run, measure and revise, and those loops only pay off when they persist, get checked, and are readable by the model doing the work.

Harness scaling is largely a feedback story. Optimizing the execution system around frozen weights lifted several models on Terminal-Bench 2.1, and the same runbook carried over to newer models unchanged, reaching 95.3% on GPT-5.6 and adding 5.4 points to DeepSeek-V4 Flash (Can execution harnesses lift model performance without retuning weights?). Running automated research loops across many environments found four reusable mechanisms that cut token traffic by 44.7–49.0% with comparable performance (Can agent harnesses be automatically optimized across many environments?). So feedback compute can buy efficiency as well as accuracy. This kind of update is also cheap. Self-improving agents split into a slow loop that changes weights and a fast loop that changes prompts, memory and tools, and recent progress clusters in the fast one because those edits are cheaper and reversible (Do self-improving agents really split into two distinct loops?).

Compare this with plain token spending. In multi-agent systems, about 80% of performance variance tracks token budget rather than coordination cleverness (How does test-time scaling work at the agent level?), and deep research agents' search steps scale much like reasoning tokens (How does test-time scaling work for individual research agents?). Undirected spending works, but feedback compute is the directed kind. On very long optimization tasks, the best predictor of success across 17 frontier models was persistence in benchmark-edit-incorporate cycles, not the quality of the first attempt. Most models quit early or burned their budget without learning anything (What predicts success in ultra-long-horizon agent tasks?). A dollar spent inside a loop that closes is worth more than a dollar spent on one longer attempt. Structure can do part of the job too: a behavior-to-code map let weaker planners match stronger models' code localization while using 8–13% fewer planner tokens (Can explicit behavior maps help weaker planners compete with stronger models?).

Feedback is effective only if the model can use it. Models at every tier are about equally good at proposing harness edits, but the benefit from those edits follows an inverted U that peaks in mid-tier models. Weak models fail to invoke the harness, and strong models struggle to follow its instructions faithfully (Do stronger models always evolve harnesses better?). The feedback also has to be verified. In large agent networks, coordination degrades because agents accept neighbors' information without checking it, so errors spread (Why do multi-agent systems fail to coordinate at scale?). Unchecked feedback adds noise rather than signal.

The payoff can extend past the harness. A deployed routing harness already records trajectories, difficulty estimates and outcomes that can be turned into labeled training data, so the serving system doubles as a data source for fine-tuning (Can a routing harness generate its own training data automatically?). The alternative shows what is at stake. Agents trained only on static expert demonstrations never see their own failures, so their competence is capped by what the curators imagined (Can agents learn beyond what their training data shows?). Feedback from real interaction is what lets an agent go beyond that ceiling, whether the update lands in the harness or in the weights.