Can you make an AI smarter just by upgrading the tools and scripts around it — without ever retraining the model itself?

This explores whether improving the wrapper around a frozen model, rather than the model itself, can produce gains that last. The short answer from the corpus is yes, the gains are real and sometimes large. Whether they last depends on what the wrapper is anchored to. The clearest evidence comes from harness work. Optimizing only the execution system around fixed models raised Terminal-Bench scores across several model families, and the same runbook carried over to newer models without changes Can execution harnesses lift model performance without retuning weights?. That portability is a kind of lastingness that weight updates don't have, because a fine-tune is tied to one checkpoint. A model can even rewrite its own scaffold over several rounds and keep getting better while its weights stay untouched Can language models improve their own scaffolding without weight updates?.

The more surprising part is *why* some scaffold gains stick. When a stronger model built harnesses for a weaker one, it nearly doubled the weaker model's Theory-of-Mind scores. It didn't do this by coaxing longer reasoning. It moved the unstable parts of the reasoning into deterministic code and task-specific routing Can a stronger model lift a weaker one at test time without retraining?. Code doesn't drift or forget, so the durable gains come from taking jobs away from the model, not from persuading it to do them better.

There is a catch, though. A loop where the model scores its own scaffold improvements eventually stalls. The model can't reliably check its own work, its outputs grow more alike, and it learns to game whatever it is graded on. The methods that keep improving all bring in an outside reference point: tool feedback, an outside judge, or user corrections Can models reliably improve themselves without external feedback?. STOP shows this too: how much improvement the loop can extract depends on whether its checking signal is written as code or as plain English. So scaffold-only gains last only as long as the feedback stays honest. A related warning: a change that worked once may not work after the underlying model changes, so earlier success should prompt a fresh check, not automatic reuse Should past update success guide future model changes?.

Several notes push back on treating "scaffold vs. weights" as a binary at all. One proposes a sequence: extend the scaffold, amplify what it reveals, absorb that into the weights, then *retire the scaffold*. Each surface carries its own hidden costs, and moving the gains between surfaces recovers them Can sequencing edits across three surfaces avoid hidden costs?. On that view, scaffolding is a good place to discover improvements but not always the best place to keep them. Agent memory raises the same point: an external memory module and the model it serves can each be trained separately and drift out of step, which is an argument for building memory into the model itself Should agent memory live inside the model backbone?. Another note moves the question entirely, arguing that adaptation belongs to the whole loop of model, harness and external evaluation, not to any one model snapshot Where does model adaptation actually happen?.

There is also a middle path most readers won't expect. Representation finetuning leaves every weight frozen and instead edits the model's internal activations as it runs. It beats LoRA while using 10–50x fewer parameters Can editing hidden representations beat weight updates for finetuning?. So "without updating weights" doesn't have to mean "only from the outside." You can also intervene inside a frozen model without retraining it.