Enabling multi-agent collaboration in knowledge graph environments

Iñaki Arango, Ayush Noori, Lucas Vittor, Joaquin Polonuer, Marinka Zitnik

Keywords: multi-agent, collaboration, knowledge graphs, git, compression, version control

TL;DR: We introduce GraphWorld, an environment for agents to build and edit graphs.

Abstract: Knowledge graphs (KGs) are critical for grounding large language models and providing them with persistent memory. However, the development of agents capable of collaboratively building and maintaining these KGs is hindered by a lack of suitable environments. Current tools often obscure the construction process, lack standardized editing APIs, and offer poor support for concurrent, multi-agent collaboration. To address this, we introduce GraphWorld, an environment for agents to build and edit graphs. GraphWorld provides agents with tools to create, update, or delete nodes, edges, or properties of KGs, with support for Python, TypeScript, and Rust. GraphWorld relies on Diamond, a compact, property-preserving storage format that permits scaling to multi-million-edge KGs. All changes are version-controlled by integrating graph-aware diffing and merging directly into Git, enabling multi-agent and human-agent collaboration through standard branching workflows.

Archival Option: The authors of this submission do *not* want it to appear in the archival proceedings.

Submission Number: 131