Interesting approach to the dynamic skill problem. I'm curious about the overhead. How much latency does the semantic search step usually add before the agent actually executes the task?