AI workflows built on agents and skills are probabilistic from end to end. The agent picks its own tools, data and models, and it can choose differently on every run. For security teams, that means credentials scoped too broadly, governance that stops at the workflow level instead of reaching the model and MCP calls inside it, and no clean audit trail. So either a human checks every output, or AI stays away from anything that matters. In this session, Barndoor AI’s Neil Mansilla and Jay Parisi will show how to build automations that isolate the steps that need AI reasoning, run the rest deterministically, and govern every model and tool call through an AI gateway. We’ll discuss: → Why patching skills to chase reliability has a ceiling → How limiting inference to the steps that need judgment makes agent behavior predictable, auditable, and brings costs down → How to enforce least-privilege access at the model and tool-call level → How to scope guardrails once, then distribute workflows safely across teams