Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve

01 ABSTRACT

To address predefined graph interfaces limiting the evidence agents can access, the paper proposes DKGI: agents write Python code directly over complete node and edge files to follow relations, combine node sets, and select textual evidence, interleaving execution with shell search and file inspection. On benchmark datasets across three domains, training-free DKGI exceeds the strongest published baseline by at least 6.7 Hit@1 points on every graph; controlled ablations show multi-turn programming outperforms shell-only access and predefined graph tools. The authors view these findings as highlighting the value of letting tool creation and tool use co-evolve during graph search.

02 KEY FINDINGS

1. DKGI lets agents create and revise graph access tools directly over complete node and edge files, rather than being constrained by predefined graph interfaces.
2. Agents write Python code to follow relations, combine node sets, and select textual evidence, interleaving this with shell search and file inspection; execution feedback guides tool creation and revision.
3. On benchmarks across three domains, training-free DKGI exceeds the strongest published baseline by at least 6.7 Hit@1 points on every graph; ablations show multi-turn programming outperforms shell-only access and predefined graph tools.