Researchers from Princeton University have demonstrated that a multi-agent system can outperform individual general-purpose models. By combining specialized agents developed by different teams, the researchers achieved approximately 70% accuracy on the SWE-bench benchmark, which evaluates software engineering and reasoning capabilities.
The proposed framework, described as a marketplace of agents, functions similarly to a hospital where specialized professionals handle specific tasks. This approach draws on concepts from ensemble learning and microservices architecture, allowing autonomous agents to collaborate on complex problems rather than relying on a single, larger model.
The study suggests that future development may shift from simply scaling model size to optimizing the coordination and composition of diverse, specialized agents. This transition introduces new challenges, including the need for systems to manage agent selection, context sharing, and output integration.
The researchers noted that while the results are significant, real-world application remains complex. Coordinating multiple agents introduces new failure points, as errors from one specialist can propagate through the chain, and identifying the source of incorrect outputs becomes more difficult compared to single-model systems.
