Microsoft Build 2026 reframes Windows as an agent platform built around three core components: Windows Agent Framework (WAF), Foundry Local GA, and Project Polaris, with Polaris set to replace GPT-4 Turbo in GitHub Copilot starting August 2026.

Foundry Local GA ships a roughly 20 MB embeddable runtime compatible with Windows, macOS Apple Silicon, and Linux x64, enabling local inference without per-token cloud charges or an Azure subscription, using an in-process architecture with no separate daemon; Azure AI Foundry GA adds a visual RAG designer, per-project token budgets, and catalog entries for Cohere, Mistral, and Stability AI.

Copilot’s coding agent remains generally available in VS Code and JetBrains, supporting the end-to-end workflow—from issue to branch to tests to PR—while leveraging security and quality tools like CodeQL, secret scanning, and dependency review, with models routed by task across Claude Sonnet 4.6, GPT-4o, and Gemini 2.0.

Reporting draws on Microsoft Build 2026 coverage, including developer blogs and related trade press.

This Windows agent framework movement sits alongside Nvidia’s Vera CPU, Microsoft MXC sandboxing, and Anthropic IPO activity, framing Microsoft’s desktop agent OS layer as a rival to broader enterprise AI model deployments.

Key takeaways: WAF with YAML manifests and a June 2026 preview, Foundry Local GA with a ~20 MB runtime and no per-token charges, Polaris poised to replace GPT-4 Turbo for Copilot by August 2026, plus Azure Agent Mesh and Agent Store for a distributed agent fabric and implications for developer workflows and Copilot model changes.

WAF lets developers define agents via YAML manifests that run on Windows 11, Windows Server 2026, Windows 365 Cloud PCs, and Azure Arc edge nodes, with a .NET/Python library inside the OS container providing connectors to Azure OpenAI, OpenAI, and the GitHub Copilot SDK; a runtime preview for text agents handling JSON, XML, and PDF rolled out to Windows Insiders in June 2026, with a 2027 roadmap for vision/UI.

Project Polaris is Microsoft’s mixture-of-experts coding model running on Maia accelerators, designed to supplant GPT-4 Turbo as the default Copilot engine starting August 2026 and reduce reliance on OpenAI inference for that path.