AI agents can be tuned with memorized, task-specific fixes, yet those help middling models most, not the weakest or strongest.

This explores why the payoff from an agent's evolved scaffolding (its 'harness': the prompts, tools and rules wrapped around a model) peaks in middle-capability models, when most of those evolved edits are memorized fixes for specific tasks. The corpus documents the two halves separately and doesn't test the link directly, so the link below is my reading of them.

The tier pattern is the surprising part. The ability to *write* useful harness edits is roughly flat across model tiers, but the ability to *benefit* from them follows an inverted U that peaks in the middle. Do stronger models always evolve harnesses better? gives a reason at each end. Weak models often fail to invoke the harness at all, so the fix exists but never gets used. Strong models struggle with faithful instruction-following. They have their own way of doing things, and the harness's rules don't reliably override it.

The second half explains why the middle is where it pays off. Analysis of evolved harnesses finds the edits are sensible but mostly memorized: they record fixes for specific tasks that an agent could have rediscovered in a single rollout (Do harness edits learn reusable strategies or memorize task fixes?). These are shortcuts to things already within reach, not new strategies. A mid-tier model sits where a cached shortcut matters. It can read and follow the note, but it wouldn't reliably land on the fix by itself every time. A weak model can't act on the note. A strong model was probably close to reaching the fix anyway, and it is the tier that follows instructions least faithfully.

The same finding puts a ceiling on the effect. Because the edits cache what's already within reach, they don't turn hard failures into successes, so even the mid-tier win is modest. It is a convenience, not a capability boost. That is also why the gains need careful measurement. How should we measure gains from automatic harness evolution? argues that a harness only earns credit for gains beyond what plain test-time search achieves with the same feedback and inference budget. If memorized fixes mostly save the agent a rollout it could have run anyway, a matched-budget baseline may shrink the mid-tier advantage.

Model tiers fail in different ways, so they need different kinds of help. A separate result on document editing shows the same thing: weaker models visibly delete content, while frontier models silently corrupt it (Does model capability change how documents degrade?). A fix written for one failure mode won't fit another, and mid-tier models are the ones whose failures a memorized patch happens to fit.