Presentation: From Retrieval to Reasoning: Building Production-Ready Agentic AI Systems with Knowledge Graphs

Details

This is a 48-minute QCon AI conference presentation from Cassie Shum, VP of Ecosystem and Product Engineering at RelationalAI, covering her production experience building and running agentic systems on knowledge graph foundations.\n\nThe four patterns she identifies: (1) Context bundling — structuring what an agent receives to optimize token usage without losing relevant relationships. (2) Decision provenance — using the knowledge graph to track why an agent made a decision, enabling audit and debugging. (3) Code as truth — treating the knowledge graph schema as the authoritative source rather than natural language documentation. (4) Agent visibility — building harness infrastructure to observe and steer agent behavior in production.\n\nShum positions knowledge graphs as moving beyond basic RAG by encoding explicit relationships, enabling agents to reason over connected data rather than retrieving semantically similar chunks. The engineering harness she describes automates feedback loops and optimizes token usage across multi-agent pipelines.\n\nThe presentation draws on 20+ years of software engineering experience and practical deployments at enterprise clients. Source is InfoQ, a reputable technical conference content platform. Content is substantive but comes from a conference talk transcript, so depth varies by section.