A small question about browsecomp benchmark's agent design 😊

[fifia07212025]

Hi everyone,

I'm trying to reproduce MiroFlow these days and got a little confused about the browsecomp benchmark. I noticed that it doesn't use sub agents like other benchmarks do — instead, it handles both reasoning and searching in the main agent.

I'm wondering what the purpose of this design is. Is it because the browsecomp task is relatively lightweight, or to avoid the complexity of multi-agent coordination?

Any insights from the team or community would be super appreciated! Thanks for building such a cool framework!