Governance is the prerequisite for capturing value from agentic AI. As agent systems proliferate, the firms that win will not be those with the most capable agents, but those with the governance infrastructure to trust, measure, and improve them. Governance operates on two layers: an input layer (constraining what the agent can see, reason over, and do) and an output layer (measuring whether the agent is actually delivering business outcomes). Effective output governance requires a structured eval system, not an ad hoc checklist. Start from strategic business outcomes, decompose the agentic workflow into its lifecycle steps, and derive eval dimensions from that decomposition. Organize those evals into a scorecard, tier your investment (Goals & Guardrails → Operational → Ad Hoc), and use a human-in-the-loop design, LLM judges to scale measurement, human experts to validate judge quality and catch drift. Governance is not the brake on velocity, it is what makes speed safe. The endgame is an AI-native product improvement loop where agents sample production failures, generate code fixes, test against the full eval suite, and ship with humans supervising rather than executing. That loop only closes when every step is governed by trusted evals and tiered quality gates.