Inside the Agent Runtime: GitHub Copilot Harness in Microsoft Copilot Studio

11:00 AM

What turns an LLM into an agent capable of completing real work? In this technical session, we’ll go inside the GitHub Copilot harness powering a new generation of agents and workflows in Microsoft Copilot Studio. We’ll examine how the harness orchestrates reasoning, context and tool execution, and how instructions, knowledge, tools, skills, memory, MCP and connected agents fit into the architecture. We’ll compare this approach with the standard Copilot Studio harness and discuss how to decide which architecture fits a particular scenario. Along the way, we’ll use practical examples to demonstrate how reasoning-heavy agents can break complex goals into steps, interact with multiple capabilities and adapt their execution as they work toward an outcome. You’ll leave with an architectural model for designing Copilot Studio agents and a better understanding of where agent development is heading.

Speakers

Dejan Pervulov

Senior Partner Solution Architect - Microsoft