I have been using sllim (note: I am the author) which allows me to easily parallelize the network requests to speed up agents.

It is only helpful with distributed chains (like tree of thoughts) and doesn't help with sequential chains.

link: https://news.ycombinator.com/item?id=36913492