How to Know When an AI Agent Can Work Unsupervised — Evan Kormos, Coinbase

Evan Kormos works at Coinbase and has spent 15 years in enterprise financial services. His talk is about bringing agents into policy-driven, regulated workflows, and this account comes from the talk's YouTube description and chapter list, not a transcript. The first chapter is titled "Policy is the original system prompt". The likely point is that in regulated businesses, written policy already did for human workers what a system prompt does for an agent, though the description does not spell that out.

From there, the description says, he maps a business process as a DAG, a directed graph of steps where work flows one way from one step to the next, and then looks inside each step at the loop where the work is done and checked. His central claim is that such a loop is only as trustworthy as its check. On that basis he presents evals (tests that score an agent's output) and traces (records of what the agent actually did) as a new confidence mechanism for organizations putting agents to work.

The practical answer to the title is a sequence of three stages. An agent first works fully supervised, then has only a sample of its work reviewed, and finally people handle only the exceptions. The chapter list says the talk also covers exception flows and bottlenecks, how input quality limits output quality, and how to choose which steps an agent should own. The description does not say what evidence or thresholds move an agent from one stage to the next, so readers who need those criteria will have to watch the talk.