How to make LangChain agents faster?

Langchain agents works not bad with multiple tools however until the final output is given, the response times are usually around 40 seconds. Has anyone tried out an architecture that made them much more faster without compromising on the output quality?