A new technique lets AI agents store their skills in model weights instead of stuffing them into every prompt.
Researchers built LatentSkill, a framework that uses a pretrained hypernetwork to convert text-based skills, the reusable task procedures agents currently paste into context windows, into LoRA adapters that load like plug-ins. Tested on the ALFWorld household-task benchmark and a Search-QA benchmark, it beat standard in-context skill prompting: ALFWorld success rose 21.4 points on seen tasks and 13.4 points on unseen ones, while prefill tokens dropped 63.9% on average. On Search-QA, exact-match accuracy rose 3.0 points while using 71.8% fewer tokens per step. The resulting skill adapters also cluster into a coherent semantic structure, can be dialed up or down with a scaling coefficient, and can be combined through simple parameter arithmetic when skills are aligned.
Context windows aren't free. Every skill description an agent pastes into a prompt adds latency and cost, and it sits in plaintext, which matters if those skills encode business logic worth keeping private. Moving skills into weights instead of tokens is part of a broader shift away from the assumption that longer context windows can absorb unlimited agent overhead.
It's a research result, not a shipped product, and a hypernetwork that generates LoRA weights on the fly introduces its own complexity worth scrutinizing before anyone bets a production agent stack on it.