v3.0 · May 2026
Designing the agent and the supervisor.
A coding agent deleted a production database in nine seconds. A procurement agent confirmed three hundred and forty transactions that never happened. A clinical decision support system ran for six to eight years at half its advertised accuracy before anyone audited it.
Each of these is the same failure mode, in different industries, on different timescales.
This book is the field manual for designing agentic products that survive the day the failure mode arrives.
Most teams shipping agentic AI design half a product. They build the agent, the autonomy boundary, the tool integrations, the runtime workflow, and treat the human side as a rollout problem. This book is for the product managers who understand that the supervisor is the second product, and that designing it is the work that separates pilots that scale from pilots that quietly get rolled back.
Three ideas run through every chapter. AI is probabilistic, not deterministic, and that changes the PM toolkit at first principles. You are designing two products at once, the agent and the human system that supervises it. And the supervisor is not a fixed input. The supervisor population is being reshaped by the same deployment that depends on it.
The regulatory and design default for AI is “keep a human in the loop.” Sustained AI use erodes the human’s ability to perform that loop’s function, in software engineering, in customer service, in clinical medicine, in every field where the supervisor’s competence depends on practice.
The book is for the product manager who needs to design supervision that does not assume the supervisor is the unimpaired version of themselves they were on day one.
Phase 0 is a permanent state of the organization, not a project stage.
AI Literacy
A state of the organization, not a project stage.
Key question
Does this PM understand what an agent is and is not?
Discover & Decide
Decide whether an agent is the right solution. Most failures prevented here.
Key question
Should we actually build this?
Design
Design the agent and the supervisory system. Runtime behavior engineered deliberately.
Key question
What can the agent do unilaterally?
Evaluate
Prove readiness through evidence. Replaces the pass/fail gate with distribution.
Key question
What is the P10 pass rate across K runs?
Observe
Measure what the agent actually does. Continuous front end of Operate.
Key question
Did the agent stay within its boundary?
Operate & Retire
Act on what Observe shows. Drift detection, governance, retirement criteria.
Key question
Is the agent still doing what we authorized?
Yoram Friedman is a physician and senior product leader who has spent fifteen years building enterprise data and AI platforms at SAP, Walmart, and Hello Heart. He trained as a physician (anesthesiology and radiology, Israel) and writes about healthcare AI, agentic systems, and what it actually takes for AI to move from pilots to production in regulated environments.