The hidden tech debt of agentic engineering

Cool agentic demos are everywhere: PR summarizers, triage agents, and vulnerability fixers. They work well on their own, but teams hit roadblocks when they try to roll them out to more users.

In production, the agent itself is only a small piece of the puzzle. Everything around it, including context, security and permissions, human-in-the-loop processes, quality control, governance, and change management, is where teams must invest most of their effort.

The speaker sees this firsthand while working with engineering teams building agentic workflows. One team using an agent is manageable. Ten means 90% of capacity goes to building infrastructure, not agents.

This talk explains what breaks at each stage of adoption and presents a framework for the infrastructure agents need to run in production.