AI Agents Fail in Production Despite Passing Evals

Despite 85% of enterprises piloting AI agents, only 5% have deployed them to production, according to Cisco data. Amazon AGI director Bryan Silverthorn told VB Transform 2026 that reliability, not capability, is the core barrier blocking enterprise deployment. Silverthorn, who joined Amazon through its Adept AI acquisition, argues reliability must be understood across four dimensions: consistency, robustness, predictability, and safety. He says these factors are routinely conflated in evaluations, causing agents to pass internal tests but fail in real-world customer environments.