How we built our multi-agent research system