RelationalAI's Cassie Shum built her team's development harness on a knowledge graph

"This is not a dogmatic, everybody needs to use a knowledge graph," Shum said ^[7]. She put the same limit on the rest of the material: "It is also not like a GraphRAG tutorial" ^[8], and "This is also not a framework prediction. I don't know what's going to happen" ^[9]. She said, "This is just Cassie's experience and how I work with knowledge graphs and the agents and everything that we're doing now" ^[18].

Shum does her knowledge graph work at RelationalAI ^[3], and she is direct about the resulting bias. "I see knowledge graphs everywhere. I see everything in graphs because that's what I've been doing for the past three years," she said ^[14]. In the demo she is her own user: "I'm the customer of a knowledge graph in this particular demo, and the lessons learned there," she said ^[13]. She describes running agentic systems as "It's like the gambler thing" ^[16], which is the most honest account of an agent loop I have read this week.

The reason she gives for the approach is about modelling. "The reason I got into graphs is because I do think modeling data and modeling your domain is very graphical in format. It's about relationships," she said ^[19]. So the modelling comes first: entities, and the links between them, before an agent queries any of it. Twenty-plus years into a career that started in code and moved into consulting at Thoughtworks and elsewhere, she went back into the codebases to see what the agents were doing, and built the harness for her team out of that ^[4][5].

The revision rate is in the record. Her QCon New York presentation, in November of last year, went much deeper into creating a knowledge graph with GraphRAG ^[10]. She has pivoted away from that on the strength of lessons from the last six months, and says everything changed again in between ^[11]. "The worst part about giving presentations is you have to change it all the time now because it's rapidly moving," she said ^[12].

For the retrieval-to-reasoning framing to transfer to another system, that system's hard cases would have to be relationship cases. That is measurable before anything gets built: count the agent queries whose answers require following two or more links between entities, then check how many of those your current retrieval already answers correctly. Shum's harness answers that question for one team and one domain ^[5].