From Business Metrics to Behavioral Personas Controllable User Simulation for Pre-Deployment Agent Testing

Pre-deployment testing of conversational agents requires simulating diverse user behaviors, yet real interaction data is scarce for new products and scripted test cases miss critical edge cases. Existing LLM-based user simulators generate personas from demographic templates or fictional backstories, lacking grounding in operational data, validated behavioral controllability, and systematic coverage of agent failure modes. We present a framework that addresses these gaps by grounding persona generation in operational business metrics, parameterizing behavior along continuous personality dimensions, and designing adversarial archetypes top-down from hypothesized failure modes via greedy set-cover. Controllability validation reveals that personality parameters produce large behavioral effects (41\% of dimension--feature pairs with Cohen's $d{>}0.8$) yet frequently non-linear responses---only 19\% show strong linear correlation---highlighting a calibration challenge for prompt-steered simulators. We demonstrate the framework as a quality gate for a sales advisory agent, where it exposes a ${\sim}$37pp goal-rate gap between stateless and state-tracking architectures ($p{<}0.001$), with per-failure-mode diagnostics that directly informed architecture selection.