An AI that keeps a fresh, shared skill library gets more capable with every run — not just every model upgrade.

This explores why keeping durable, regularly-updated working artifacts makes an agent's harness compound its capability instead of stalling — and the corpus suggests the answer is about where knowledge lives and how it's maintained, not raw model strength. The core mechanism is externalization: when hard-won know-how is written into a persistent store the agent consults, each run can build on the last without overwriting weights. VOYAGER shows this most cleanly — executable skills kept in an embedding-indexed library, with complex skills composed from simpler ones, let an agent learn continuously while dodging the catastrophic forgetting that weight-update methods suffer Can agents learn new skills without forgetting old ones?. The artifact is the memory; the model stays frozen and free.

But 'persistent' alone isn't enough — the resynchronization matters, because artifacts rot. The ACE framework's key move is treating context as an evolving playbook updated through generation-reflection-curation loops rather than full rewrites, precisely to prevent brevity bias and 'context collapse,' where compression quietly erases the details that made the artifact useful Can context playbooks prevent knowledge loss during iteration?. Why that discipline is non-negotiable becomes vivid in the failure case: frontier models silently corrupt about 25% of document content across long delegated relay chains, with errors compounding round after round and never plateauing Do frontier LLMs silently corrupt documents in long workflows?. Left unmanaged, artifacts compound decay just as readily as capability. Structured, incremental resync is what flips the sign.

The compounding is strongest when curation is treated as its own job. SkillOS decouples a trainable curator from a frozen executor, and the repository drifts away from generic verbose additions toward actionable execution logic and reusable cross-task meta-strategies — and that trained curator generalizes across different executor backbones Can a separate trained curator improve skill libraries better than frozen agents?. This is the deep reason artifacts beat raw scale: the capacity to *produce* useful harness edits is roughly flat across model tiers, while the capacity to *benefit* from them peaks in the middle — weak models never invoke the harness, very strong ones over-follow it Do stronger models always evolve harnesses better?. Capability gains come from the artifact ecosystem, not from a bigger brain reading it.

Structure in the artifact is what lets gains stack instead of blur. MetaGPT finds that standardized engineering documents outperform conversational coordination, with agents pulling from a shared environment rather than passing noisy messages Does structured artifact sharing outperform conversational coordination? — the same logic scales down to a single agent's harness. And when the persistent layer holds governance rather than skills, the effect repeats: an agent that recorded 889 governance events over 96 days followed rules encoded in the memory it actually consulted during decisions, which beat external policy precisely because it was resident in the loop Can governance rules embedded in runtime memory actually protect autonomous agents?.

The quietly surprising takeaway: the bottleneck on compounding isn't storage or retrieval fidelity — it's how work gets broken into artifact-sized pieces. Decomposition granularity gates everything downstream, with standard LLM decomposition hitting only 34% step-level recall and correcting the step count recovering 75% of the gains What blocks skill retrieval in task decomposition?. A related move, reconstructing memory by actively traversing and pruning a graph rather than fetching fixed chunks, buys up to 23% on reasoning tasks Can agents reconstruct memory on demand instead of retrieving it?. Persistent resynchronized artifacts compound because they turn one-shot cleverness into a maintained, composable asset — but only when something disciplined keeps re-cutting them to the right size and pruning the rot.