Can an AI that rewrites its own scaffolding learn reusable strategies, or does it only memorize one-off shortcuts?

This explores whether harness evolution can be redirected from memorizing task fixes toward distilling transferable procedures. The starting problem is real and well-documented: when you inspect what an evolving harness actually writes, most edits are rational but shallow — they cache a shortcut for a task the agent could already have solved in a single rollout, rather than converting a failure into a repeatable strategy Do harness edits learn strategies or just memorize fixes?. That's exactly the behavior 'distilling transferable procedures' would replace, so the question is whether the loop can be re-pointed.

The first thing the corpus says is: prove the redirection is real before celebrating it. Harness evolution is itself a search process, so any 'better procedure' it claims to have found is confounded with raw search effort — only gains that survive a matched compute budget and hold up on tasks the harness never saw during evolution count as genuine, transferable design Are harness evolution gains really from better design?. Transfer, in other words, is defined by held-out performance, not by how sensible the edits look.

On the how, the most promising lever is representation, not more editing. Reorganizing a harness around runtime *behavior* — an explicit map from what the system does to the code that does it — let weaker planners match much stronger models at localizing the right code, while using fewer tokens Can explicit behavior maps help weaker planners compete with stronger models?. That's the shape of a transferable procedure: a durable structure that carries capability across tasks rather than a patch bolted onto one. The same instinct shows up in how skills get shared — aggregating interaction trajectories from many users and running an evolver that extracts recurring patterns turns siloed one-off fixes into collective, reusable skills How can agent systems share learned skills across users?. Distillation of procedures seems to need cross-task or cross-user pressure; a single agent optimizing its own scaffold on its own tasks has no reason to generalize.

Here's the part you might not expect: even if you succeed, distillation may not raise the ceiling. On-policy distillation — the closest analog to 'compressing a teacher's procedure into a student' — turns out to catalyze exploration within the student's existing capability envelope rather than expand it, and the quality and diversity of the guiding signal matters far more than the size of the source Does on-policy distillation actually expand student capability?. Redirecting harness evolution toward transferable procedures would likely make agents *reliably reach* what they can already do, not do genuinely new things. And who benefits is uneven: the capacity to produce useful harness edits is flat across model tiers, but the capacity to actually benefit from them peaks in the middle — weak models don't invoke the procedures, strong ones already have them internalized Do stronger models always evolve harnesses better?.

So the redirection is plausible but constrained. Techniques exist to make procedures stable and portable — a proximal teacher held inside a trust region converts unstable distillation into monotonic improvement Can proximity between teacher and student fix distillation instability?, and adapters treated as persistent behavioral state show how a learned delta can be reused at scale rather than re-derived each time Can lightweight adapters replace millions of personalized models?. The catch the corpus keeps returning to: verify transfer on held-out work, expect exploration gains rather than new capability, and aim the pressure at mid-tier agents who can actually cash the procedures in.