Brian Kennedy

Brian Kennedy is a principal AI software engineer at ZoomInfo, where he leads engineering for the Agentic AI Platform behind Agent Teams. By Brian’s count, the platform powers more than a hundred production agents. One decision his team faces is when to let a workflow advance. If a grader approves a step too soon, the next step can depend on work the agent never did.

Brian spent a decade at early-stage startups, most recently building generative AI products and platforms. He worked as a software engineer and technical lead at Future, a personalized remote coaching company. In 2023 he joined Woven as a founding staff software engineer. His team combined language models and agents in an AI travel partner that helped travelers discover places to go and plan itineraries. He then spent a year as a member of technical staff at Writer before joining ZoomInfo in December 2025.

How a workflow decides to move on

At ZoomInfo, Brian works on the runtime that directs agents across a customer’s business systems. Inside GTM Studio, Agent Teams uses business signals and ZoomInfo data to coordinate that work. Brian has spent roughly a year on the orchestrator that directs each step, the completion checks that let the workflow advance, and the evaluations that test those checks.

Agent Teams used a second model to grade each step an agent took. Brian and his colleagues ran Jev, a model that answers structured questions about the information it receives, alongside that grader. The comparison uncovered cases where the grader had approved work the agent never performed. That finding shapes Brian’s argument in his upcoming CODE talk, Your Judge Is Hallucinating Too. Before judging how well an agent performed an action, the system has to establish whether the action happened.