Post by Yubin Kim

When and Why do we want to use Multi-Agent System (MAS) over Single-Agent System (SAS)? 🤖 🤖🤖 vs 🤖 Excited to share our research at Google Research & DeepMind: "Towards a Science of Scaling Agent Systems" As agents become the dominant paradigm for real-world AI, the community is shifting from demonstrations to design principles. A persistent question remains: "How do we move from heuristic choices to principled agent system design?" To answer this, we conducted 180 controlled experiments across • 5 agent architectures (SAS and four MAS variants) • 3 LLM families (OpenAI, Google, Anthropic) • 4 diverse, agentic benchmarks (Browsecomp-Plus, WorkBench, Finance Agent, Plancraft) and derived quantitative scaling principles for agent systems. Key insights: 📊 Architecture-Task Alignment > Number of Agents Simply adding more agents rarely helps. Matching coordination structure to task properties matters far more. 🔬 Three effects we identified: 1) Under fixed compute budgets, tool-heavy tasks are more sensitive to multi-agent overhead. 2) When single-agent baselines exceed ~45%, coordination yields diminishing returns. 3) Orchestrator-Workers MAS contains error amplification to 4.4× vs 17.2× in independent systems. 📈 Our model (R²=0.513) predicts optimal coordination strategies for 87% of held-out configurations based on measurable task properties. 📄 Paper: https://lnkd.in/e325up7S Deeply grateful to my amazing collaborators Ken Gu, Ali Heydari, PhD, Zhihan Zhang and mentors Yilun Du, Samuel Schmidgall, Chunjong Park, Xuhai "Orson" Xu, Yuzhe Yang, Tim Althoff, Daniel McDuff and Xin Liu for their guidance throughout this journey. #Agents #MultiAgentSystem #AgentScaling