AI brokers want abilities — particular procedural data — to carry out duties effectively, however they’ll’t train themselves, a brand new analysis suggests.
The authors of the analysis have developed a brand new benchmark, SkillsBench, which evaluates agentic AI efficiency on 84 duties throughout 11 domains together with healthcare, manufacturing, cybersecurity and software program engineering. The researchers checked out every activity beneath three situations: no abilities (agent receives directions solely), with curated abilities (supplied with a listing, code snippets and sources to assist it) and self-generated abilities (agent has no abilities however is prompted to develop them).
Typical duties included conducting a safety audit of npm dependencies for vulnerabilities, or analyzing differential protein expression in most cancers cell line knowledge.
