Patterns and problems in emerging multi-agent systems

As AI agents increasingly perform tasks in shared systems, their interactions are expected to surge, potentially surpassing human levels. Early models like Sonnet 4.6 and Opus 4.6 struggled with coordination, often creating conflicting code contributions. Newer models, including Opus 4.8 and Mythos Preview, mitigate conflicts by limiting shared work. However, Sonnet 5 successfully operates on shared resources while maintaining high throughput. A key concern is that individual agents' "low variability" might lead them to make similar erroneous decisions, causing systemic failures rather than isolated issues.