For coding agents, real-time collaboration beats the "wisdom of the crowd" (ykdojo.github.io)

Recent experiments with AI coding agents, specifically Gemini 3.5 Flash, have revealed that real-time collaboration significantly enhances problem-solving accuracy in programming tasks. When tested on 30 Project Euler problems, five agents working together in real-time achieved an average accuracy of 87%, compared to just 72% when operating independently. Additionally, the majority voting method, often likened to the "wisdom of the crowd," provided only marginal improvement, with accuracy rising from 80% to 90%. These findings underscore the potential for collaborative AI systems to outperform traditional individual approaches. The study also highlighted notable efficiency gains from collaboration. Agents working in tandem not only produced higher accuracy but did so in less time, with many collaborative runs completing within the first five minutes. This contrasts sharply with solo agents, which can sometimes take nearly 40 minutes per problem due to overlap in working time. This research has significant implications for the AI/ML community, suggesting that developing interactive multi-agent systems, like xAI’s Grok Heavy, could pave the way for more effective and efficient AI problem solvers. Overall, these insights point to a promising direction for the future of AI collaboration.