Key Takeaways from Microsoft Build 2025 on AI, Copilot, and NPUs

Written By:
May 22, 2025

Microsoft Build 2025 kicked off with a clear focus: empowering developers to build smarter, faster, and more securely with AI. From renaming Copilot Runtime to Windows AI Foundry to introducing native support for multi-agent systems and Model Context Protocol (MCP), Microsoft is creating a unified AI development stack, local-first, hardware-accelerated, and deeply integrated into Windows. Whether you're deploying models across NPUs or building agentic apps, this year’s Build is packed with updates that directly impact how developers design, optimize, and scale AI-driven software. Here's a breakdown of the most relevant announcements:

Developer-Centric Takeaways:

1. Windows AI Foundry replaces Copilot Runtime with a unified local AI platform.
The renaming of Windows Copilot Runtime to Windows AI Foundry is more than cosmetic. This rebrand aligns Windows with Azure AI Foundry, offering parity in lifecycle support, from model selection and optimization to fine-tuning and deployment, directly on-device. For developers building latency-sensitive or privacy-preserving AI features, this creates a local-first stack capable of bypassing the cloud entirely.

2. Windows ML remains the core inferencing layer, with multi-silicon acceleration.
At the core of Foundry is Windows ML, the inferencing engine that supports ONNX models and runs efficiently across a wide range of hardware:

  • CPU and GPU (Intel, AMD, NVIDIA)

  • NPU acceleration (Qualcomm Hexagon, Intel Movidius, etc.)

This flexibility means developers can bring their own models (BYOM) and still take advantage of platform-optimized inference without needing to rewrite for each hardware backend.

3. Foundry Local integrates model catalogs like Ollama and NVIDIA NIMs.
With native access to open-source model catalogs, developers can now search, preview, and deploy foundation models (e.g., LLMs, vision models) directly within local apps. This makes it feasible to:

  • Rapidly test model performance across workloads

  • Avoid custom integration overhead

  • Tailor models to local context without cloud latency
4. New On-Device AI APIs + Fine-Tuning with LoRA
  • Inbox AI APIs on Copilot+ PCs now expose ready-to-use functions for:

    • Text summarization, OCR, image captioning, object removal

    • Custom prompt workflows for dynamic UI interactions

  • LoRA fine-tuning is supported natively for the Phi Silica SLM (Small Language Model), enabling developers to personalize models with their own data, on-device, without retraining from scratch.

  • New APIs for Semantic Search & RAG:

    • Developers can now build semantic search and retrieval-augmented generation pipelines with local data directly inside Windows apps, no external cloud dependency needed.
5. Model Context Protocol (MCP): Standardizing Agent-App Interactions
  • Windows 11 adds native support for MCP, allowing apps to act as context providers or skill endpoints for local AI agents.

  • Think of MCP as a universal adapter: Agents can now dynamically interface with any app exposing MCP hooks (like commands, views, data), enabling a modular agentic system.

  • Microsoft is rolling MCP into:

    • GitHub Copilot Studio, Semantic Kernel, Azure AI Foundry, Dynamics 365, and now Windows itself.
6. App Actions: Declarative Entry Points for App Functionality
  • App Actions allow developers to expose granular app functions (like “generate report,” “start timer,” or “analyze image”) as invokable actions at the OS level.

  • These actions are discoverable by Copilot, agents, and the system, offering deeper OS integration and new user acquisition channels.

  • Designed with intent-based invocation in mind, App Actions let your features surface contextually across Windows, reducing friction and boosting usage without modifying core app logic.
7. New Developer-Facing Security APIs: VBS Enclave SDK + PQC
  • The VBS Enclave SDK enables apps to execute sensitive operations in isolated, hardware-backed enclaves, leveraging Hyper-V-based Virtualization Based Security.

  • This is ideal for scenarios involving credential handling, DRM, or confidential inference workloads, keeping code and data shielded from the OS or other processes.

  • Microsoft is also introducing Post-Quantum Cryptography (PQC) APIs to future-proof apps against quantum threats, allowing devs to start adopting quantum-safe encryption algorithms today.
8. Windows Subsystem for Linux (WSL) Is Now Open Source
  • Microsoft is officially open-sourcing WSL, including key components like:

    • The WSL kernel

    • WSLg (GUI support)

    • The Mariner-based distro backend

  • Developers now get direct access to the WSL internals and APIs, paving the way for:

    • Custom Linux integration layers on Windows

    • Deep tooling extensions

    • More community-driven feature development (e.g., sandboxed dev environments or WSL-native agent hosting)

  • Also announced: The GitHub Copilot Chat extension for VSCode will be open-sourced soon, enabling developers to customize their own AI workflows inside the editor.
9. Dev Tooling Upgrades: Terminal, WinGet, PowerToys
  • Windows Terminal now supports tab tear-out, pane management improvements, and deeper customization.

  • WinGet adds support for package pinning, configuration files, and bulk installs, streamlining dev machine setup and CI/CD environments.

  • PowerToys introduces new modules that extend automation and clipboard workflows, making it even more appealing for power devs scripting repetitive UI tasks.
10. Microsoft Store: Dev-Focused Distribution Enhancements
  • New free developer registration removes the paywall for publishing apps to the Microsoft Store.

  • Web Installer for Win32 apps allows one-click installs directly from the browser, no MSIX packaging required.

  • Expanded analytics and App Campaign programs offer visibility and growth tooling to help developers track usage, optimize engagement, and run store-driven promotions.
11. Multi-Agent Systems with Copilot Studio (Preview)
  • Microsoft is making multi-agent architectures accessible via Copilot Studio, enabling developers to design systems where multiple AI agents collaborate by delegating and coordinating tasks autonomously.

  • These multi-agent systems can be built leveraging:

    • Microsoft 365 Agent Builder

    • Azure AI Agents Service

    • Azure Fabric

  • This unlocks new possibilities for complex workflows, such as multi-step automation, distributed problem-solving, or composable AI services that mirror human team dynamics.

  • Currently available in preview, Copilot Studio invites developers to experiment with agent orchestration at scale.
12. NLWeb: Transforming Websites into Agentic Apps
  • Natural Language Web (NLWeb) is a Microsoft initiative to convert traditional websites into agent-driven, conversational interfaces.

  • By enabling natural language queries and interactions directly against web content, NLWeb empowers developers to create agentic web apps that are more intuitive and context-aware.

  • While OpenAI’s direct involvement is unclear, NLWeb signals a push toward embedding conversational AI deeper into everyday web experiences.
13. GitHub Copilot Evolves: From Pair Programmer to Peer Programmer
  • The new GitHub Coding Agent (“Project Padawan”) is now generally available for Copilot Enterprise and Pro+ users.

  • Unlike the classic Copilot, Project Padawan is designed to autonomously handle low-to-medium complexity tasks within well-tested codebases, including:

    • Adding new features

    • Bug fixing

    • Test extensions

    • Refactoring

    • Documentation improvements

  • This shift represents a move toward AI as a true collaborative peer, capable of more independent coding contributions rather than just suggestions.
14. Copilot Studio Gains “Pro” Developer Capabilities
  • Microsoft continues to evolve Copilot Studio beyond its low-code/no-code origins into a robust platform for professional developers building complex AI agents.

  • New APIs for Microsoft 365 Copilot are being introduced, with the retrieval API currently in preview, allowing devs to integrate intelligent data retrieval into their agents.

  • Importantly, Bring Your Own Model (BYOM) support from Azure AI Foundry is now in preview for Copilot Studio, enabling developers to deploy custom-trained models seamlessly within agent workflows, a major step toward flexible, enterprise-grade AI solutions.
15. Integrated AI Security with Entra, Defender, and Purview
  • Microsoft embeds security tools directly into Azure AI Foundry and Copilot Studio to protect AI applications throughout development and deployment.

  • Entra Agent ID (now in preview) provides identity and access management tailored for AI agents, supporting integrations across Microsoft tools and third-party platforms like ServiceNow and Workday.

  • This signals Microsoft’s push to ensure AI agent governance, compliance, and security are baked into the AI dev lifecycle, not bolted on afterward.

Microsoft Build 2025 made it clear: AI is now a local-first, developer-first priority. With Windows AI Foundry unifying the model lifecycle, Copilot Studio enabling collaborative agents, and expanded hardware acceleration, Microsoft is delivering a full-stack AI platform built for real-world development.

Add in stronger security, open-sourced tools like WSL, and seamless deployment across platforms, developers now have everything they need to build smarter, faster, and more securely. The future of software is AI-native, and Build 2025 just handed developers the blueprint.

Connect with Us