Tag
This paper examines why skills in LLM agents work by stabilizing execution through procedural anchoring, while also identifying limitations like retrieval bottlenecks and brittle assumptions.