Microsoft used its Build 2026 keynote to ship Windows Agent Framework 1.0, an MIT-licensed open source toolkit for building AI agents that operate inside Windows the way a human user would. The release elevates a long-running preview to a production 1.0 and folds it into the Windows App SDK as a first-party component.
The framework provides a unified runtime for tool calling, structured screen reading, file-system scoping, permission prompts, audit logging, and inter-agent message passing. It also exposes Windows-native primitives — UIAutomation trees, accessibility APIs, the clipboard, and Win32 shells — to model code through typed schemas, so an agent can click a button, read a confirmation dialog, or paste a value without scraping pixels.
Crucially for enterprise adoption, every agent runs inside a per-task sandbox with explicit user consent for any action that touches files, network, or other applications. Microsoft is positioning the framework as the desktop counterpart to its Azure Agent Mesh announcement, which handles federated multi-agent execution across clouds and devices.
The MIT license is the headline policy choice. By choosing one of the most permissive licenses available, Microsoft signals it wants third-party IDEs, automation vendors, and rival agent platforms (including those from Anthropic, OpenAI, and open source projects) to build directly on the Windows agent surface. The repository ships with reference adapters for OpenAI, Anthropic, Google, and Microsoft's own MAI models, and the project is hosted under the official microsoft GitHub organization.
For developers, the immediate impact is that desktop-agent prototypes no longer need bespoke screen-reading or process-isolation code on Windows — those layers are now part of the OS SDK. For users, the change is more subtle: any agent built on the framework will route every privileged action through the same standardized consent UI.
Source: [BuildFastWithAI](https://www.buildfastwithai.com/blogs/ai-news-today-june-2-2026)