Multi-agent search method for generating harder AI tool-use tests

A researcher says evaluating AI tool use is expensive and domain-dependent, and that experts need to design new tasks when test data become saturated. Their paper describes generating challenging tasks through multi-agent evolutionary search. A coauthor says the approach can raise the proportion of hard tasks on AppWorld from 19.8% to 52.5%, while making tasks more semantically and functionally diverse and keeping them realistic.