A new paper proposes speeding up how AI agents learn new tasks by having them borrow from other agents' instruction sets instead of starting from zero.
The researchers describe agent skills as reusable, natural-language instructions that guide an agent through a task on a given harness, the software environment it operates in. Millions of these skills are already shared publicly, covering many domains and harnesses, but tuning one for a new task has typically meant running the agent repeatedly through costly trial and error, ignoring all that accumulated work. The paper's method, Retrieval-Augmented Skill Optimization (RASO), instead retrieves relevant skills from an external library and adapts them to the target task and harness, correcting for mismatches in both domain and framework. It runs in two stages: one builds an initial skill from retrieved knowledge with no agent runs required, and a second refines that skill using retrieved knowledge plus feedback from actual executions. Across four agent benchmarks and two models, RASO outperformed versions of the method that skipped retrieval.
The real news here isn't the algorithm, it's the implicit claim that public skill libraries have grown large enough to function as training data, not just handy templates. If retrieval genuinely reduces the number of expensive rollouts needed to tune an agent, that lowers the cost of building capable agents for teams that can't afford endless trial and error.
Cross-harness adaptation is doing a lot of work in that sentence, and RASO's gains are measured against non-retrieval baselines rather than against skills a human expert tuned by hand, so how much this actually saves in practice is still an open question.