LLM-based multi-agent systems need effective collaboration topologies, but prior work often uses fixed granularity (individual agents or predefined groups) for all subtasks. This paper introduces MAGIC, a framework that mixes granularity locally: for each functional role, it decides whether to use a single agent or a reusable group. MAGIC constructs such graphs incrementally using dense-reward , directly optimizing the construction policy with rewards shaped by probe-based utility and structural signals.
In tests across eight benchmarks, MAGIC outperforms state-of-the-art topology generators and also demonstrates strong inference efficiency. By learning to balance fine-grained control and reusable patterns within one organization, it adapts to varying subtask complexities while keeping execution costs manageable. This makes it a practical step toward building more capable and efficient multi-agent systems.