From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills

Researchers evaluated model-generated skills across five agent task domains using a utility-based framework.

The skills helped on average, but their benefit varied widely and sometimes caused negative transfer.

Extractor and consumer behaviors were uneven, showing that skill quality depends on how experiences are generated and used.

The authors proposed a meta-skill that improves extracted skill quality and reduces harmful transfers.