Should an AI agent grow a shared code library over time, or patch its scaffolding fresh for each new task?

This explores whether agents should invest in durable, shared code artifacts rather than patching their harness anew for each task — and the corpus's honest first answer is that we don't fully know yet, because the durable-and-shared case is the least studied part of the whole system. One survey of agentic code elements finds that agent-initiated artifacts that persist and get shared across agents are exactly the underexplored layer, with the open questions clustering around lifecycle (when to keep something), shared-state consistency, and how scratch work gets promoted into durable infrastructure What makes agent-authored code worth persisting and sharing?. So the interesting move is to read the adjacent evidence for what tips the balance.

Where the evidence gets specific, it favors durability — but with a twist about *what* you make durable. VOYAGER shows that an externalized, reusable skill library (executable skills you index and compose later) lets an agent keep learning without catastrophically forgetting, precisely the failure that hits approaches which overwrite themselves each round Can agents learn new skills without forgetting old ones?. But raw accumulation isn't the win. SkillOS finds that a *frozen* agent left to grow its own repository tends to pile up generic, verbose additions; separating out a trained curator shifts the library toward actionable execution logic and reusable cross-task meta-strategies — and that curator generalizes across different executor backbones Can a separate trained curator improve skill libraries better than frozen agents?. The lesson: durable artifacts beat per-task patches only when something is actively deciding what deserves to persist.

There's also a structural argument for shared code over scattered edits. Reorganizing a harness repository around runtime behavior — an explicit behavior-to-code map — let *weaker* planners match stronger models' code localization while spending fewer tokens Can explicit behavior maps help weaker planners compete with stronger models?. A durable shared representation, in other words, can substitute for raw model strength. The Darwin Gödel Machine pushes this furthest: it keeps an evolutionary archive of agent variants and improves by empirical benchmarking rather than one-off fixes, and the capabilities it discovers are things like better context management and editing — durable infrastructure, not disposable patches Can AI systems improve themselves through trial and error?.

Here's the twist you probably didn't come looking for: the value of harness edits isn't flat across models — it's an inverted U. One study found every model tier can *produce* useful harness updates, but the capacity to actually *benefit* peaks at mid-tier models; weak models never invoke the harness they wrote, and strong models struggle to follow their own instructions faithfully Do stronger models always evolve harnesses better?. That reframes the whole question. "Durable vs. per-task" isn't a universal ranking — durable shared artifacts pay off most for the models in the middle that can both write and use them, and the promotion-from-scratch-to-infrastructure decision is the real lever.

Worth noting the cousin domains that make the same trade in different vocabulary. PEFT reframes lightweight adapters as *persistent local state* — durable behavioral deltas that let one base model plus millions of small adapters replace millions of full models Can lightweight adapters replace millions of personalized models?, while core-parameter isolation shows that naively merging many task-specific updates causes interference unless you explicitly protect what should stay stable Can isolating task-specific parameters prevent multi-task fine-tuning interference?. Same shape as the harness question at the weights level: durable shared state wins, but only with an explicit discipline for deciding what persists and keeping it from corrupting everything else.