Traditional workflows are built around explicit states, transitions, and rules. Agents are built around probabilistic reasoning and tool use. Trying to combine them forces a hard question: how much autonomy can a business process tolerate before it stops being trustworthy?

This talk looks at agentic workflows as a domain design problem. We’ll explore where deterministic workflow steps still matter, where an LLM can safely operate inside a process, and how policies, approvals, and guardrails keep the overall system predictable even when one participant is not. The goal is to find the right balance between procedural control and agent autonomy, so teams can automate work that was previously too ambiguous or high-touch for traditional workflows.