Why AI Agents Need Knowledge Graphs, Not Just Data

Most companies adopting AI agents will eventually need to manage thousands of them, and few have infrastructure built for this scale. That is the problem Eric Broda, founder of Broda Group Software, lays out with host Kris Peeters, CEO of Dataminded.

Eric Broda has spent 40 years in the industry, building ecosystems around APIs in what was first service mesh, then data mesh, and now what he calls the agentic mesh. Agents become participants in a business process, and running thousands of them needs the same discipline that data mesh gave to data products. They need ownership, governance, and a way to scale.

Document ingestion is the easy entry point. A harder case is a client using agents to track R&D tax credit exposure in real time, reviewing GitHub commits and Jira tickets as they happen instead of waiting for a year-end retrospective.

Broda also shares how his assumptions changed. Knowledge graphs work better with coarse markdown files than fine-grained data. Retrieval-augmented generation breaks down past 50,000 records. And testing agents looks surprisingly more like an employee review than like a unit test: we check if the task was completed, not if the output exactly matches the requirement.

He draws a clear line of using AI agents and autonomous systems in domains where judgment built from experience still needs a human to make the call (like healthcare diagnosis or military). He closes with a lesson from his career: relationships, not technical skill alone, drive success.