Which version controls? Context graphs give legal agents the answer 📄 A single credit agreement can come with several drafts, amendments and side letters. Before an AI agent can answer a question about it, it has to establish which version governs. Piecing that structure together from raw files is hard. Context graphs are a promising way to improve agent quality and efficiency, so the Harvey team tested the impact on Legal Agent Bench. The setup: For each task, the agent first reviewed every document in the client matter It built a graph of relevant entities and relationships: parties, dates, draft vs. executed status, and which versions belong to the same contract The same agent was then graded with and without the graph, using LAB’s standard setup The results: Up to 30% higher rubric pass rate Up to about 50% lower task cost Some cases showed much larger gains. In one firm-knowledge task, the agent had to find the firm’s most recent secured financing and retrieve the executed credit and security agreements as precedents for a new facility. Without the graph, it searched files one at a time across many matters and chose the wrong precedent: 0% of rubric criteria. With the graph, it located the relevant matter and its executed agreements immediately: 75%. Questions worth considering: is the cost of building the graph counted within task cost, and how do graph costs change as people interact with the matter over time? By Niko Grupen Harvey Graph sbee.link/ajkpnvwmf7 #ContextGraph #LegalTech #AIAgents #AgentEvaluation #LegalAI -- Connected Data London 2026 | 11–12 November | Leonardo Royal Hotel London Tower Bridge 🎟 Tickets on sale now - 🔔 Full programme announced sbee.link/wj8xdfayg6 📺 Sponsorship opportunities available. Contact [email protected] for details. #KnowledgeGraph #GraphRAG #Ontology #Graph #AI #DataScience #GraphDB #SemTech