Towards a science of scaling agent systems: When and why agent systems work
4 days ago
- #Multi-Agent Systems
- #Scaling Principles
- #AI Agents
- Multi-agent coordination improves performance on parallelizable tasks but degrades it on sequential ones.
- Agentic tasks require sustained multi-step interactions, iterative information gathering, and adaptive strategy refinement.
- Five canonical architectures were evaluated: single-agent (SAS), independent, centralized, decentralized, and hybrid.
- Centralized coordination improved performance by 80.9% on parallelizable tasks like financial reasoning.
- Multi-agent variants degraded performance by 39-70% on sequential tasks like planning.
- Independent multi-agent systems amplified errors by 17.2x, while centralized systems contained amplification to 4.4x.
- A predictive model was developed to identify the optimal architecture for 87% of unseen tasks.
- The model uses task properties like tool count and decomposability to predict the best coordination strategy.