When Cloud Agents Meet Device Agents: Lessons from Hybrid Multi-Agent Systems

Researchers tested hybrid multi-agent AI systems that combine cloud LLMs with on-device SLMs.

The results show these setups can balance accuracy, cost, and energy better than single-model approaches in some cases.

But the best architecture depends on the task, and using larger frontier models does not always improve performance.

The study highlights a task-specific tradeoff space rather than one universal optimal design.