Multi-agent collaboration reasoning