Nothing says that agents can't read skills, interpret them and fetch stuff. Needing to agent to pass things off to powerful LLMs to interpret doesn't have to be done for everything.