Agentic Graph Learning (AGL) leverages a large language model (LLM) powered agent that sequentially samples graph evidence to support its final prediction. Existing approaches either rely on a single agent or orchestrate multiple role‑based agents over the whole graph, but both share a common reasoning policy, which is suboptimal for graphs with heterogeneous structural and semantic patterns.
To overcome this limitation, we introduce MAAGL (Multi‑Agent Agentic Graph Learning). MAAGL first partitions the graph into communities and assigns an independent agent to each community, enabling region‑specific specialization. Structural and semantic evidence are handled separately: structural evidence is summarized by a structural signature, a dynamically updated, fixed‑size representation that is permutation‑invariant; semantic evidence is filtered to the top‑$k$ most relevant nodes after each sampling step.
Agents estimate their confidence based on historical trajectories with similar signatures. When confidence falls below a threshold, a debate‑style collaboration is triggered: agents share their structural signatures and semantic summaries, cross‑question each other, and fuse their viewpoints, thereby improving overall reasoning quality.
Extensive experiments on four benchmark datasets demonstrate that MAAGL consistently outperforms state‑of‑the‑art AGL methods across accuracy, recall, and other metrics, validating the effectiveness of community partitioning, structural signatures, and collaborative debate.
Review: MAAGL preserves graph permutation invariance via structural signatures while exploiting multi‑agent specialization and collaboration, successfully breaking the bottleneck of a single reasoning policy and offering a promising direction for large‑scale graph reasoning.