By ByteBot
Microsoft announced the Windows Agent Store at Build 2026, and the economics are worth pausing on: developers keep 85% of revenue. Apple and Google take 30%. Whether this generosity lasts is a fair question, but for now, it is a meaningful invitation to build. The store runs on the Windows Agent Framework (WAF), an MIT-licensed runtime that lets agents live natively on Windows — not just in a browser tab or serverless function. That is a different kind of surface area, and it comes with 1.4 billion potential users on day one.
Two Stores, One Platform
Microsoft drew a distinction at Build between two agent distribution channels. The Windows Agent Store is developer-facing: you publish a manifest, pass a security review, set a price, and your agent is listed alongside apps in the Microsoft Store. The Windows Agent Gallery is consumer-facing — a UI panel in Windows where home users discover free and premium agents, each tagged with privacy labels modeled on Apple’s Nutrition Labels. Agents handling sensitive data — financial records, health information — must declare it. That requirement exists because agents operate with ambient permissions that most apps never request.
The YAML Manifest Is the Core Primitive
Every agent in the Windows Agent Framework starts with a declarative YAML manifest. The manifest describes the agent’s capabilities, intents, permitted actions, and safety constraints. The runtime reads it — you do not bake deployment targets into code. The same manifest that runs as a local process on your development machine can escalate to a Windows 365 GPU node under load, or deploy as an Azure service, without re-architecture.
Visual Studio 2026 (v17.12) ships an Agent Designer that generates YAML manifests through a low-code interface, along with project templates that deploy to a sandboxed local agent runtime on F5. If you prefer writing the manifest by hand, the schema is published and version-controlled like any other config file.
How to Build and Publish an Agent
Publishing to the Windows Agent Store requires four steps:
- Define your manifest. Use the Visual Studio Agent Designer or write YAML directly. Declare capabilities, intents, actions, and safety constraints.
- Build and test locally. Install the Microsoft Agent Framework —
pip install agent-frameworkfor Python,dotnet add package Microsoft.Agents.AIfor .NET — then run it against the local agent runtime. - Create a Partner Center account and enroll in the Microsoft AI Cloud Partner Program (MAICPP). This is the same Partner Center used for Windows apps.
- Submit and pass the security review. Microsoft reviews every agent before listing. The review checks for over-broad permissions, data exfiltration vectors, and manifest accuracy. Expect a stricter bar than the standard app review — agents have broader system access by design.
Once listed, the 85% revenue share applies to both one-time purchases and subscriptions. Free agents are listed without a review fee.
What Is Available Now
The Windows Agent Runtime preview reaches Windows Insiders in June 2026. In this release, agents can operate on structured text data — JSON, XML, and PDF files. Vision-based agents that process screen pixels are on the roadmap for 2027, pending optimization of the underlying on-device multimodal model.
For enterprise-scale deployments, Azure Agent Mesh enters preview alongside the store. It is a control plane that routes agent tasks across on-premises Windows servers, Windows 365 Cloud PCs, and Azure Arc-enabled edge devices based on latency and GPU availability. General availability is targeted for Q4 2026 with consumption-based pricing.
AutoGen and Semantic Kernel Are Done
Microsoft Agent Framework 1.0 reached general availability on April 7, 2026. It merges AutoGen and Semantic Kernel — two frameworks that accumulated more than 50,000 combined GitHub stars and an increasingly confused developer community — into a single production SDK. Both predecessors moved to maintenance mode: bug fixes only, no new features. Migration from Semantic Kernel typically takes 2–4 hours for an average app. Migration from AutoGen is harder because the programming model shifted from conversation-centric to graph-based.
The framework supports six model providers out of the box — Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, and Ollama — with provider switching as a one-line configuration change. MCP tool integration and Agent-to-Agent protocol are native at 1.0, not afterthoughts.
The Honest Take
The 85% revenue share is not charity. Microsoft lost ground to Claude Code in developer adoption, and the Agent Store is partly a bid to win that back by making Windows the most economically attractive platform for agent developers. The agent runtime’s current limitation to text-based structured data means vision-heavy agents are not yet viable Store products. And the security review gate is real — expect it to catch anything that smells like over-privileged access.
That said, if you are already building agents with Agent Framework or its predecessors, publishing to the Store is a logical next step. The same YAML manifest you run locally is what gets listed. The Agent Framework GitHub repository has functional starting points, and if you are migrating from Semantic Kernel or AutoGen, the Visual Studio Magazine breakdown of Agent Framework 1.0 is the clearest external overview available. The migration guide on the Agent Framework Dev Blog covers the path in detail.
