Descriptions:
Phil Hetzel, head of solutions engineering at Braintrust—an agent quality platform built around evals and observability—makes a deliberately provocative argument: agentic AI development does not naturally belong to data scientists or ML engineers, even though that is where most traditional enterprises have routed it.
Drawing on 12 years of consulting (including leading the global Databricks business unit at Slalom) and a year at Braintrust watching agent teams across many industries, Hetzel identifies two distinct organizational archetypes. Traditional enterprises delegate GenAI to existing ML platform teams because “AI is in the name,” while AI-native startups build small cross-functional engineering teams from scratch. The core problem with the traditional route, he argues, is that data scientists apply the wrong success metrics—precision, recall, F1—to problems that require much broader functional evaluation of agent behavior in the real world.
The countercase Hetzel builds is that LLMs are fundamentally APIs, and product engineers who compose APIs daily may be closer to the user problem and better positioned to evaluate whether an agent is actually working. He stops short of a definitive prescription, instead framing the talk as a prompt for organizations to audit how they have structured ownership and whether inherited ML processes are the right fit for a fundamentally different technology. For engineering leaders and CTOs currently deciding how to staff agentic projects, the structured for/against framing makes this a useful reference point.
📺 Source: AI Engineer · Published May 25, 2026
🏷️ Format: Opinion Editorial
