### Session Room C · 10:45 AM–11:25 AM ### Where Agents Break: Deterministic Governance for Probabilistic AI Every agent looks brilliant in testing. Most fail in production — not because the model is wrong, but because nobody built the tracks. This session walks through a live, running governance framework built on top of CrewAI that proves one thesis: reasoning can be agentic, scoring must be computational, and execution must be governed. We'll cover the three surfaces where agents actually break — human trust, compliance boundaries, and downstream propagation — and show exactly how a deterministic NumPy scoring layer, a per-agent CLEAR handoff system, and a three-model judge array physically block unsafe execution before it reaches production. This isn't theory. We'll run the demo live: watch the terminal output, see the governance checkpoint fire mid-flow, watch autonomy get denied by math — not by a model's self-assessment. You'll leave with a framework-agnostic governance layer you can drop into any CrewAI, LangGraph, or raw Python flow today.  **Preston McCauley**Clear Sight Designs, LLC Preston McCauley is a cognitive systems architect and AI Innovation Leader with 25 years of experience shipping real products across generative AI, agentic systems, augmented reality, healthcare diagnostics, and computational biology. He is the founder of Clear Sight Designs, LLC, where he helps teams design and execute AI strategy with a single principle at the center: understand where it fails before you build for success. He developed the CLEAR AI governance framework after observing the same pattern across dozens of production systems — the model wasn't the problem. The process was. His work sits at the intersection of deterministic and probabilistic systems. He is the architect of PHYSIM — a deterministic computational drug-discovery engine that maps disease architecture across 6,675 conditions, achieving 97.8% clinical concordance. No training data. No confidence scores. Same molecule, same answer, every time. The engine validates drug candidates through a 7-gate biological gauntlet, replacing years of pre-clinical screening with weeks of computation. That same architectural philosophy — isolating what must be deterministic from what can be probabilistic — is what drives his governance work in agentic AI. The CLEAR governance layer he will demonstrate live at AgentCon is built on the same principle: reasoning can be agentic, scoring must be computational, and execution must be governed. He is the author of Generative AI for Everyone, a practical guide to applied AI literacy, and is completing his second book, Leading with an AI-First Mindset. In 2025, he was nominated as one of the 75 AI Innovators in Dallas.