The Agentic AI Team

The team ships the agent. Then it discovers the agent was only half the product, and no one owns the other half.

Your last book taught you that every agentic product is two products: the agent that acts, and the supervisory layer that decides what it may do, when it must stop and ask, and what happens when it is wrong. This book is about who builds that second product, because it turns out the team you have was built for a different one.

For two decades, a software team had three seats: a product manager, a designer, an engineer. That triad was the right operating model for software a human drives. An agentic product is not that. It decides, it acts, it drifts after it ships, and the work of supervising it does not land cleanly on any of the three seats. It falls into the seams between them, and the seams are where agentic products fail.

This is the book about the team the supervisory layer actually requires. It is written for the experienced product manager who now has to build a team around them that the old playbook never named: an architect who turns a boundary into a wall the agent cannot cross, an eval owner who can say whether the agent is still correct and not just still running, an agent supervisor whose actual job is to watch the thing in production, a domain expert who holds the interest of the person the agent affects and is never in the room. Most teams have not created these roles. The responsibilities exist whether the titles do or not, and on most teams the honest answer to "who owns this one" is, for at least two of them, no one.

The failures in this book are real, and each one lives in a seam: the agent that deleted a production database in nine seconds through a boundary nobody enforced, the eval that passed while the product was quietly wrong, the gate that gave a human about a second per decision, the model that degraded for months while every dashboard stayed green. None of them is one person's failure. That is the point.

Inside, you'll learn:

• Why the PM/designer/engineer triad is no longer sufficient, and the second pipeline it leaves unstaffed
• The collaboration grid: four kinds of work across two channels, eight cells, and the column the old model never filled
• The new owners an agentic product forces into being, the architect, the eval owner, the agent supervisor, the context owner, the domain expert, and what each one holds
• Why silent degradation is invisible to any single owner: six drift vectors, six sightlines, and no one who can see all six
• How the team's deliverables change shape, the user story becomes a graded outcome spec, the journey map becomes a boundary map, and each new artifact pulls in an owner it never had
• How to staff the agentic team at a startup, a scale-up, and an enterprise, and the supervision seat the salary market has not learned to price yet
• What happens when agents start building agents, and the supervision problem recurses one level up
• Why the line a human can never hand to an agent is accountability, and what that means for every org chart

This is not a book about what agents are. It is a book about the team that has to own one, and the half of that team most companies have not hired.

The author is a physician and senior product leader who has shipped enterprise software at Microsoft, SAP, and Walmart and built regulated health technology where being wrong has consequences. The examples come from high-stakes environments, healthcare, finance, operations, where a wrong autonomous action is not a bug ticket but a consequence someone absorbs.

If you are responsible for an agentic product, or about to be, this is the book about the team you have not built yet.