D2 Product Preview: Agentic Safety, Semantic Authorization & Guardrails for Real-World AI Actions

As more teams move beyond chat over data and start giving agentic systems the ability to take real actions—triggering workflows, changing state, even touching money—the old safety patterns start to crack. Role-based access control on APIs and a few regex-based guardrails on model output aren’t enough when you have autonomous or tool-using agents making decisions inside complex, ambiguous business processes. Agentic systems today are substantially limited by hallucination, adherence problems, poor context length, increasingly high token use, and model attention.

In this roundtable, we’ll explore how to make agents significantly more effective, using emerging patterns for agentic safety: how we direct and constrain what agents are allowed to do, on whose behalf, and under what conditions. We’ll focus on semantic authorization—authorizing “what this action means in the business context” rather than just “which endpoint is called”—and on practical, layered guardrails that go beyond prompt engineering or AI red-teaming.

Chad Wathington is the CEO of Manual Software. His career has spanned product, strategy, and engineering leadership roles. Some of you may have attended his session @ ELC Annual 2024. He will give a brief overview and demo of a semantic authorization control plane, and then the group will spend most of the time comparing notes:

1. How are you preventing unsafe or unintended agent actions today?
2. Where do current tools and patterns fall short for your domain?
3. What would a credible safety, governance, and audit story look like for agents in your organization?