It reuses previously generated code, so tools it creates persists from session to session. It also lets the LLM avoid actually “seeing” the tokens in some cases since it can pipe directly between tools/write to disk instead of getting returned into the LLMs context window.