Researchers have built a filter that helps AI agents tell apart skills that look nearly identical on paper but behave very differently in practice.
The method, called SkillContrast, is a training-free selector for AI agents choosing which "skill" or tool definition to use for a task. Instead of just matching a query to skill descriptions, it compares retrieved skills against each other and keeps only the text that differs between them, feeding that distinctive context to the reranker rather than entire skill documents. On a 1,235-request benchmark called SameCapRisk-Bench, built from skill pairs that do similar things but carry different risk levels, SkillContrast produced 54 to 72 more "clean hits" than a TF-IDF baseline, meaning cases where the agent retrieved a useful skill without also pulling in its riskier lookalike. That range comes from testing four combinations of two retrievers and two reranker sizes, not from one inconsistent result. Against feeding rerankers the full skill text, the new approach cut model-input tokens by 51.1 to 58.8 percent, though a smaller 0.6-billion-parameter reranker saw 10 to 18 fewer clean hits under the leaner inputs, while a larger 4-billion-parameter reranker matched or beat the full-text baseline on clean hits.
This is a narrow fix, but it points at a real blind spot in agent tool retrieval: similar skills often share boilerplate instructions and differ only in the few lines that determine whether a tool is safe to use, and plain query matching tends to blur that distinction. As more companies wire tool and skill libraries into autonomous agents, picking the right instruction text is quietly becoming as important as picking the right model.
It is still an arXiv paper tested on one lab-built benchmark, not a shipped product, so the token and accuracy gains are promising lab numbers, not a settled answer to agent tool-selection risk.