MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning
MAGIC builds multi-agent collaboration graphs by choosing, role by role, whether to use a single agent or a reusable group.
The paper argues that existing topology generators are too uniform, using either individual agents or predefined groups across the whole system. MAGIC instead constructs a mixed-granularity graph incrementally, selecting roles, instantiating them, and wiring them into the organization. Its reinforcement learning setup adds dense intermediate feedback from probe-based utility and structural signals while preserving the final task reward. The authors report gains over state-of-the-art baselines on eight benchmarks, with an efficiency study showing strong inference performance. ArXiv · AI/CL/LG's note
The paper argues that existing topology generators are too uniform, using either individual agents or predefined groups across the whole system. MAGIC instead constructs a mixed-granularity graph incrementally, selecting roles, instantiating them, and wiring them into the organization. Its reinforcement learning setup adds dense intermediate feedback from probe-based utility and structural signals while preserving the final task reward. The authors report gains over state-of-the-art baselines on eight benchmarks, with an efficiency study showing strong inference performance. ArXiv · AI/CL/LG's note
score 4