No comments yet.
I built this to address what I see as the fundamental problem with ReAct-style agents: compounding errors. Even a small mistake made early enough in the loop can snowball and ruin the final output. But with search, agents can look multiple steps ahead and backtrack before committing to a particular trajectory. This has already been shown in a few papers to help agents avoid mistakes and boost overall task performance, but there's no easy way to actually build these kinds of agents. So that's why I made this framework. I believe search will eventually become table stakes for building agents as inference gets faster and cheaper, and this library is the first (and only) way to get that performance boost easily.
ā Jonathan
