Current agent evaluations predominantly reward what gets done (end states, correct API calls) while overlooking how it is done. This outcome-first view not only creates a misleading picture of performance but also hides harmful behaviors. An agent might “succeed” at a technical task yet fail the user with hallucination, policy violation, or sociolinguistic failure. Similarly, simple success rates mask the danger of failures: A safe handoff to a human is treated the same as a costly mistake, like issuing a refund without permission. We propose a framework to answer these questions holistically and systematically. We identify 11 consistency relationships between what agents observe, what they do, and what they tell the customer, and formalize procedure contracts as typed constraints over these relationships: what must never happen (Forbid), what must precede what (Order), what communications must be grounded in retrieved data (Ground), and what falls within authorized scope (Scope). When we applied this framework to frontier models on realistic service scenarios, we found 27–78% of benchmark successes are corrupt: correct outcome, broken process. Each model fails differently, and smarter agents make smarter mistakes that are harder to catch before they reach your customers.