Microsoft Build 2026: Windows Agent Framework Goes Open Source, Project Polaris Takes Aim at GitHub Copilot's Future
Microsoft open-sourced its Windows Agent Framework, unveiled Project Polaris to replace GPT-4 Turbo in Copilot, and shipped Copilot Workspace as GA at Build 2026.
Microsoft Build 2026 ran June 2 and 3 at Fort Mason Center in San Francisco, and Satya Nadella opened the keynote with a clear statement of direction: Windows is no longer a platform built exclusively for human users, and AI agents are now first-class citizens across Microsoft’s developer stack. Three announcements defined the conference and each one has direct, practical consequences if you build software or run teams that do.
The Windows Agent Framework Is Now Open Source
The Windows Agent Framework (WAF) v1.0 shipped under an MIT licence at Build. It is Microsoft’s library for building agents that run across local Windows machines, Windows 365 Cloud PCs, and Azure Arc-enabled edge devices, and it has been in development since Microsoft merged the foundations of Semantic Kernel and AutoGen into a single SDK last October.
The practical design decision worth understanding: agents are defined in YAML, not tied to a specific runtime. That means an agent can start as a local process on a developer’s laptop, scale up to a Windows 365 GPU node when the task demands more compute, and publish to Azure as a managed service, all from the same manifest file. No re-architecting for each environment.
Under the hood, WAF 1.0 ships with an Agent Registration Service (a local daemon managing agent health and versioning), a Cross-Agent Communication Bus built on gRPC for pub/sub messaging between agents, and a Memory Service that stores conversational context, user preferences, and learned patterns in an encrypted, user-controllable cache. The graph-based workflow engine supports sequential, concurrent, handoff, and group chat orchestration patterns, with checkpointing so long-running processes survive interruptions.
For teams currently writing their own agent infrastructure, the MIT licence means you can adopt, fork, or embed this without licensing friction. The stable 1.0 APIs also come with a long-term support commitment, which matters if you are building on top of it for production systems.
Project Polaris: Microsoft Builds Its Own Coding Model
This is the announcement with the most immediate commercial weight. Project Polaris is Microsoft’s in-house AI coding model, and it will replace GPT-4 Turbo as the default model for GitHub Copilot subscribers starting August 2026.
The context matters here. GitHub Copilot has 4.7 million paid subscribers, but it has been losing ground to Cursor and, more recently, Anthropic’s Claude Code. Microsoft’s response is to take control of the full stack: the model, the inference infrastructure (running on Microsoft’s own Maia AI accelerators inside Azure), and the developer experience. The dependency on OpenAI’s GPT-4 Turbo as the default is ending.
Polaris uses a mixture-of-experts architecture with sub-modules tuned for specific programming languages and frameworks. Microsoft says it outperforms GPT-4 Turbo on HumanEval and MBPP benchmarks, with particular improvements in lower-adoption languages like Rust and Haskell. Pro tier subscribers get multi-file context up to 100,000 lines and autonomous test generation. Chain-of-thought and tree-of-thought reasoning at inference time allow it to handle complex multi-file refactoring. One demonstrated capability was autonomously migrating a legacy .NET Framework application to .NET 9, including dependency resolution, code modernisation, and CI/CD pipeline updates.
What this means for your team specifically: the migration to Polaris is automatic in August. Microsoft is offering a three-month fallback period for teams that want to stay on GPT-4 Turbo while they evaluate the change. If you are using GitHub Copilot in production workflows or building on top of the Copilot SDK, you should use that window. Test your specific use cases against Polaris behaviour before it becomes the default, rather than discovering differences after the fact.
Copilot Workspace Is Now Generally Available
Copilot Workspace exited beta and went GA at Build. If you have not looked at it since it launched in preview, it has developed considerably.
Workspace gives Copilot a context window across an entire repository, not just the file you have open. From that context, it can propose multi-file edits, run tests, interpret the results, and iterate on scoped tasks autonomously. The GA release adds integrations with Jira, Datadog, and ServiceNow, callable directly from within an active workspace session, and a Fleet mode where Copilot CLI can work autonomously on narrowly defined codebase tasks without requiring per-step developer confirmation.
The significance of shipping this as GA alongside the agent framework is deliberate. Microsoft is positioning Workspace as the developer-facing front end of a broader agent platform. The underlying infrastructure is WAF; Workspace is where developers interact with agents in the context of actual code.
The Bigger Picture
The thread connecting all three announcements is the same: Microsoft is shifting AI from a feature that responds to prompts to infrastructure that runs work. That is the stated framing from Nadella’s keynote and it is reflected in the release timing. WAF reaching 1.0, Polaris replacing a third-party model as the Copilot default, Workspace going GA — these are not separate product launches so much as different layers of the same stack becoming production-ready at the same time.
For developers, the practical upshot is that a lot of infrastructure you might have been building yourself is now available as open-source or managed services with stable APIs. The MIT-licensed WAF in particular removes a significant amount of plumbing work for teams building multi-agent systems on Windows or Azure.
For teams using GitHub Copilot, August is the date to have on your radar. The model change is coming and the fallback window is finite. Test early.
The full agent framework documentation and the open-source repository are both live now if you want to get into the specifics before the summer migration lands.