Build 2026: Windows Agent Framework GA, Foundry Local, Polaris Copilot
Abhishek Gautam10 min readQuick summary
Microsoft Build June 2-3, 2026: MIT-licensed Windows Agent Framework, ~20MB Foundry Local runtime (no per-token cloud), Project Polaris MoE replaces GPT-4 Turbo in Copilot August 2026.
Read next
- GitHub Copilot Token Billing Live June 1: AI Credits, Dev ReactionGitHub Copilot switched to token-based GitHub AI Credits on June 1, 2026. Pro still $10 with $10 credits; devs praise fairness vs premium requests. Code review uses Actions too.
- Microsoft MXC: OS Kernel Sandbox for AI Agents — OpenAI, Nvidia OnboardAt Build 2026 Microsoft launched MXC (Execution Containers): kernel-enforced agent boundaries on Windows. Partners: OpenAI, Nvidia OpenShell, Manus, Nous Research, OpenClaw.
Microsoft Build 2026 ( June 2–3, San Francisco) repositioned Windows as an agent platform — not just a Copilot sidebar — with Windows Agent Framework (MIT license), Foundry Local GA, and Project Polaris, a homegrown mixture-of-experts model replacing GPT-4 Turbo in GitHub Copilot starting August 2026.
Windows Agent Framework (WAF)
WAF lets developers define agents in YAML manifests that run across:
- Windows 11
- Windows Server 2026
- Windows 365 Cloud PCs
- Azure Arc edge nodes via Azure Agent Mesh (consumption SKU, Q4 2026 GA cited in trade coverage)
Microsoft Agent Framework 1.0 (.NET / Python, graph orchestration) is the library inside the OS container — connectors for Azure OpenAI, OpenAI, GitHub Copilot SDK.
Windows Agent Runtime preview hits Windows Insiders in June 2026 — initial scope: text agents on JSON, XML, PDF; vision/UI agents roadmap 2027.
Foundry Local: ~20 MB, No Per-Token Bill
Foundry Local reached GA as an embeddable ~20 MB runtime on Windows, macOS Apple Silicon, and Linux x64:
- No Azure subscription required for local inference path (per keynote summaries)
- No per-token cloud charge for on-device runs
- In-process — no separate daemon
Azure AI Foundry GA adds visual RAG designer, per-project token budgets, and catalog entries for Cohere, Mistral, Stability AI.
Local-first teams should compare spend with LLM API Pricing before assuming cloud defaults.
Project Polaris and Copilot
Polaris is Microsoft's MoE coding model on Maia accelerators — replacing GPT-4 Turbo as default GitHub Copilot engine August 2026, reducing OpenAI inference dependency for that path.
Copilot coding agent is GA in VS Code / JetBrains: issue → branch → tests → PR, with CodeQL, secret scan, dependency review — models routed across Claude Sonnet 4.6, GPT-4o, Gemini 2.0 by task.
How This Fits Nvidia and OpenAI News Same Week
- Nvidia Vera CPU — datacenter agent orchestration (explainer)
- Microsoft MXC — Windows kernel sandbox (explainer)
- Anthropic IPO filing — enterprise Claude scale (explainer)
Microsoft is claiming the desktop agent OS layer; Nvidia claims factory silicon; Anthropic claims enterprise model revenue.
Key Takeaways
- Build 2026: Windows Agent Framework (MIT), Agent Runtime preview June 2026
- Foundry Local GA: ~20 MB local runtime, no per-token cloud charge for on-device path
- Project Polaris: Copilot default August 2026 on Maia, displacing GPT-4 Turbo for that route
- Azure Agent Mesh + Agent Store — distributed agent fabric narrative
- For developers: YAML agent manifests, local Foundry eval loops, plan for August Copilot model change
Sources
FAQ
Frequently Asked Questions
What is the Windows Agent Framework announced at Build 2026?
The Windows Agent Framework is an MIT-licensed platform for defining AI agents in YAML manifests that run across Windows 11, Windows Server 2026, Windows 365 Cloud PCs, and Azure Arc devices, integrated with Microsoft Agent Framework 1.0 for orchestration.
What is Microsoft Foundry Local?
Foundry Local is a roughly 20 MB embeddable AI runtime Microsoft announced at general availability during Build 2026. It runs on Windows, macOS Apple Silicon, and Linux x64 without requiring an Azure subscription for local inference and without per-token cloud charges for on-device execution.
When will Project Polaris replace GPT-4 Turbo in GitHub Copilot?
Microsoft announced Project Polaris, an in-house mixture-of-experts coding model on Maia accelerators, will replace GPT-4 Turbo as the default engine in GitHub Copilot starting August 2026.
What is available in the Windows Agent Runtime preview?
The Windows Agent Runtime preview ships to Windows Insiders in June 2026 with an initial focus on text-based agents working with structured formats such as JSON, XML, and PDF. Vision and UI-interaction agents are on the roadmap for 2027.
Free Weekly Briefing
The AI & Dev Briefing
One honest email a week — what actually matters in AI and software engineering. No noise, no sponsored content. Read by developers across 30+ countries.
No spam. Unsubscribe anytime.
More on Developer Tools
All posts →GitHub Copilot Token Billing Live June 1: AI Credits, Dev Reaction
GitHub Copilot switched to token-based GitHub AI Credits on June 1, 2026. Pro still $10 with $10 credits; devs praise fairness vs premium requests. Code review uses Actions too.
Microsoft MXC: OS Kernel Sandbox for AI Agents — OpenAI, Nvidia Onboard
At Build 2026 Microsoft launched MXC (Execution Containers): kernel-enforced agent boundaries on Windows. Partners: OpenAI, Nvidia OpenShell, Manus, Nous Research, OpenClaw.
WWDC 2026: Cook's Last Keynote — Siri Rebuilt on Gemini, iOS 27 Beta Live
Apple's June 8 WWDC keynote rebuilt Siri on a custom Google Gemini model, previewed homeOS, dropped six OS betas, and opened third-party AI extensions — two years after the original Siri promise.
VS Code Agents Go Stable Offline — Air-Gapped BYOK for Regulated Dev Teams
VS Code 1.122 brought air-gapped Bring Your Own Key agent coding — no tokens leave the network. Enterprise-managed plugins landed June 5 for Copilot Business orgs.
Written by
Software Engineer based in Delhi, India. Writes about AI models, semiconductor supply chains, and tech geopolitics — covering the intersection of infrastructure and global events. 952+ posts cited by ChatGPT, Perplexity, and Gemini. Read in 167 countries.