The AI models that benefit most from saved shortcuts aren't the strongest ones — they're the ones in the middle.

This explores why memorized harness shortcuts — the task-specific fixes an agent writes into its own scaffolding — help mid-tier models most, rather than helping the strongest models most. The cleanest answer in the corpus is that two different abilities pull in opposite directions, and mid-tier models sit at the sweet spot where both are satisfied. The capacity to *produce* a useful harness edit turns out to be roughly flat across model tiers, but the capacity to *benefit* from those edits follows an inverted-U: weak models never reliably invoke the harness in the first place, while the strongest models struggle to follow the cached instruction faithfully once it's there Do stronger models always evolve harnesses better?. Mid-tier models are competent enough to trigger the shortcut and obedient enough to just run it.

That picture gets sharper once you look at what these edits actually are. Inspection of evolved harnesses shows most edits don't distill a transferable strategy — they cache a fix for a task the agent could already solve, persisting a shortcut rather than converting a failure into a success evolved-harness-edits-mostly-memorize-task-specific-fixes-rather-than-distillin. A memorized shortcut is a crutch, and crutches help exactly the class of model that can walk but not sprint. The strongest models have less to gain because they can rediscover the fix in a single rollout; the weakest can't use the crutch at all.

There's a parallel result on the planning side that reinforces this: giving a weaker planner an explicit behavior-to-code map let it match stronger models on localization while spending fewer tokens Can explicit behavior maps help weaker planners compete with stronger models?. Externalized, memorized structure substitutes for raw capability — it lifts the middle of the distribution toward the top, but the top was already there on its own. The same substitution logic explains why the benefit curve bends downward at the high end.

A cautionary thread runs underneath all of this. Because harness evolution is itself a search loop, apparent gains are confounded with sheer search effort and can amount to memorizing the test set unless you hold the search budget fixed and evaluate on held-out tasks How should we measure gains from automatic harness evolution?. Some of the mid-tier 'benefit' may be shortcut-caching that doesn't generalize — which rhymes with a broader failure mode where models reinforced on the wrong signal learn degenerate shortcuts (answer repetition, computation-skipping) that actively contaminate real capability Do overly hard RLVR samples actually harm model capabilities?.

The thing you might not have expected to learn: the reason the strongest models benefit *less* isn't that they're too good for shortcuts — it's an instruction-following problem. A memorized harness edit is a rigid output-format prescription, and there's evidence that what such prescriptions transfer is knowledge of the output space, not task understanding Does instruction tuning teach task understanding or output format?. Strong models, with richer priors of their own, are more prone to deviate from the cached format than to slavishly execute it — so the very sophistication that lets them solve tasks unaided is what makes them worse at cashing in someone else's memorized answer.