Most “AI features” still behave like a smarter search box. Real value starts when systems can plan, act, explain themselves, and improve over time without turning your environment into a compliance horror story.

In this session we’ll design an agentic architecture that fits the way Power Platform solutions are designed. You’ll see how to move from single turn prompts to multi step execution with Copilot Studio, how to ground agents in enterprise data, and how to give them tools that are safe by default. We’ll cover memory and context, when to persist it, when to avoid it, and how to keep user intent in charge.

Then we go practical: evaluation and regression testing for prompts and actions, telemetry that explains why the agent made a call, and ALM patterns that ensure agent behavior is versioned, reviewed, and deployable across environments. The goal is simple: agents that can do real work, explain what they did, and get better release after release.