GitHub’s AI agents and Copilot Fusion

What happens when AI agents can build, review and improve software across the whole development process? At GitHub Universe 2026, Kim Isenberg speaks with Jay Parikh, Microsoft’s Executive Vice President of CoreAI, whose organization includes GitHub. Parikh explains how Copilot Fusion picks and combines different models for each task, why that is harder than a simple router, and how teams choose between keeping a human in the loop and letting agents work with a human on the loop. He describes Copilot code review, which has delivered 200 million findings since its launch, and Copilot Dreaming, which spots outdated agent instructions and suggests fixes. The conversation also covers developer choice of models and tools, compliance in regulated industries, and why Parikh sees writing software as the way to turn the potential intelligence stored in AI models into real-world impact.