Small Agent Teams Beat the Monolith

🔥 What happened Junlin Wang's team shows that sampling many small agent teams in parallel beats both independent agents and one giant team. With a 27B model and 100 agents, ten-agent teams won 10 of 12 long-horizon tasks. 💡 Why it matters At equal dollar cost, those parallel teams beat a frontier model on half the tasks. The math: individual agents suffer low mean scores, one big team gets no variance gain — small teams balance both. This rewrites how you spend inference compute. ⚡ Our take Betting on one mega-agent is burning money. The future is swarms of small teams — cheaper, more robust, and actually scalable.