Most CS teams adopted AI agents the way they adopt software: switch it on, see what happens, and hope someone notices if it goes wrong. The result is agents with fuzzy jobs, unclear limits and no one accountable for whether they’re working. This session takes a different approach. Drawing on how Vinova runs her AI agents, it treats each agent the way you’d treat a new hire. Every agent gets a defined seat and a clear definition of good before day one. Its onboarding is built around three questions: what it can see, what it can change, and what it has to prove before it earns more responsibility. And when the seat or the workflow underneath it isn’t working, it gets a clean exit rather than a slow decline nobody wants to own. After a short walkthrough of the model with real examples, attendees will work through it on their own team’s workflows, and leave with a seat description drafted for their first (or next) AI coworker. Key takeaways: - How to write a seat description for an AI agent: scope, owner, success measures and what “good” looks like before it goes live - A practical framework for agent permissions based on what it can see, what it can change and what it has to prove - How to tell when an agent, or the workflow it sits on, isn’t working, and how to retire it cleanly - A completed first draft of a seat description for an agent on your own team