Microsoft Build 2026: the enterprise agent platform arrives

A single integrated system spanning the Microsoft stack: build an agent in GitHub, deploy it to Microsoft Foundry, and optimise it automatically with the models best suited to the job — secured and governed natively.

• Agents follow a real software lifecycle: source, test, deploy, observe, improve.
• Entra, Purview, Defender and Agent 365 make governance native to the platform.
• Every agent action generates signal — trajectories, outcomes, feedback — that is captured and fed back to improve agents under human oversight.

Satya Nadella and Microsoft leaders framed the developer story: how teams build with AI across Microsoft platforms today, a focused set of tier-1 announcements, and the customer scenarios behind them.

• More than 2,600 developers in person, with 100K+ joining online worldwide.
• A new real-time Build Live blog tracked each announcement as it went live.

The workplace-intelligence layer for agents: it captures how work actually happens across Microsoft 365, organisational systems and external sources — people, emails, documents, meetings and how they connect.

• Built on four components: Chat, Context, Tools and Workspaces.
• Work IQ APIs are generally available on June 16, on a consumption model independent of individual user licensing.

A shared semantic foundation over structured business data, giving agents a consistent understanding of enterprise data modelled in Microsoft Fabric.

Ties the IQ family together and enables retrieval planning across both enterprise knowledge and the live web. Foundry IQ Serverless + Grounding enters public preview with new knowledge sources including Work IQ and Fabric IQ.

A new, AI-first web-search stack for grounding agents in the real-time web: model-agnostic, MCP-native and publisher-first, returning relevant passages at roughly 2.5× the speed of the next best alternative.

Microsoft AI's first reasoning model: a 35-billion active-parameter model with a 256K context window, built for high efficiency at low token cost, and open in private preview on Foundry.

• Trained from scratch with zero distillation on clean, commercially-licensed data.
• On blind tests independent raters prefer it to Sonnet 4.6, and it matches Opus 4.6 on coding (SWE-Bench Pro).
• Tuned for complex multi-step instructions, long-context reasoning and code generation.

A full multimodal stack from Microsoft's own models, landing on Foundry with market-leading quality per dollar.

• MAI-Image-2.5 serves both text-to-image and image-to-image (top-ranked on the Arena leaderboard); live in PowerPoint and rolling out on OneDrive.
• MAI-Transcribe-1.5 combines state-of-the-art accuracy across 43 languages, with streaming coming soon.
• MAI-Voice-2 adds 15+ languages and new voices; MAI-Code-1 (Flash & Pro) is tuned for GitHub Copilot and VS Code.

A new approach that applies reinforcement learning inside your compliance boundary, using your own data, processes and conventions, so agents learn how your business actually works — and your custom models stay in your environment.

• Reinforcement-learning environments act as private "training gyms" for your models.
• A tuned MAI model for Excel matched GPT-5.4 while being up to 10× more efficient.
• Available today via Forward Deployed Engineering, with Copilot Studio and Foundry availability coming.

For the first time developers can tune the weights of MAI models themselves, and the models will be widely available beyond Foundry.

• MAI models also available on Fireworks AI, Baseten and OpenRouter.
• Co-designed with Microsoft's own Maia 200 silicon, already showing a 1.4× efficiency boost.
• Microsoft and Mayo Clinic are co-creating a frontier healthcare model, owned by Mayo Clinic and later offered via Azure Foundry.

Deploy and manage agents without infrastructure overhead: instant-on sandboxes per session, isolated execution, persistent memory and elastic scale — "the primitive for agents the way containers were for cloud-native apps".

Low-latency, high-throughput open-model inference with Foundry's enterprise governance and Azure data residency, regardless of the model you choose.

Routes GPU workloads to available capacity across global regions automatically, giving global reach with zero infrastructure overhead.

Bring agents built on the Microsoft Agent Framework, LangGraph, GitHub Copilot SDK, Claude Agent SDK or a custom harness — with an optimised model router balancing quality, speed and cost per agent.

• Observability and traces make agent behaviour measurable.
• Foundry continuously tunes models, harnesses, IQs, tools and actions as agents operate.

A single control plane to observe, govern and secure every agent in the organisation. Every agent shows up in one catalog — whether built in Foundry or elsewhere — with full visibility of who deployed it, what it can access, how it behaves and what it costs.

• The A365 Agent Registry surfaces unmanaged agents across endpoints.
• Extends to more than 20 types of local agents across your estate.
• ME5 is now a prerequisite; Agent 365 is included in the new Microsoft 365 E7 SKU and available standalone.

Integrate any agent with enterprise-grade identity, security, observability and governed access to Microsoft 365 data — making custom agents production-ready and composable with Agent 365.

An open, end-to-end trust stack for AI agents on any framework, anchored by two open-source projects.

• ASSERT — policy-driven safety evaluation.
• Agent Control Specification — standardises where and how to apply controls in the agent loop.

A multi-model agentic security system that deploys 100+ specialised agents to find exploitable bugs — reasoning about data flow, business logic and exploit chains — with context-aware fixes delivered directly in the Defender Portal.

A native desktop experience that brings agentic development to a much wider audience: start from an idea, an issue or a PR, orchestrate multiple agent sessions in parallel, and keep changes moving through review, CI and merge. Each session uses git worktrees so work stays separated.

Build custom agentic experiences on the Copilot runtime — tool calling, streaming, file operations, sessions, permissions and observability — now GA across Node.js/TypeScript, Python, Go, .NET and, newly, Java and Rust.

Code review becomes an adaptable agentic system, extensible via custom Skills and MCP-connected context (service catalogs, incident tools, third-party platforms). A new medium tier (preview) routes per-repo to a higher-reasoning model.

• Now also available natively in Azure Repos / Azure DevOps (public preview) with inline comments and committable fixes.
• Admins can dial review depth per repository under usage-based billing.

Terminal-native agentic workflows get isolated local sandboxes, recurring background tasks and a cross-surface memory layer.

• GitHub Sandbox — cross-platform local sandboxing for Copilot CLI and SDK with IT-governed controls.
• /every — automate recurring prompts and tasks in the background so repetitive work doesn't interrupt flow.
• Chronicle — extends /chronicle into memory that carries context across CLI, VS Code, GitHub Desktop, github.com and JetBrains.

An open-source GitHub Copilot CLI skill that connects your local project to the live Build 2026 session catalog, mapping your dependencies to relevant sessions from the terminal.

• Reads package.json, requirements.txt, .csproj, go.mod and other manifests, maps them to Microsoft products and searches the live catalog — no stale data.
• Plan a schedule, log session notes, get next steps, and scaffold a project from a session, all from the CLI.
• Install via /plugin install microsoft/Build-CLI; also works in VS Code, Visual Studio 2026 and Claude Code, with Microsoft Learn MCP fallback for SDK docs.

A policy-driven execution layer where developers declare what an agent can access (files, network) with containment enforced by the OS itself, across a spectrum of isolation semantics composable by intent and risk.

• Fast process isolation (adopted by GitHub Copilot CLI) and session isolation separate the agent from the desktop, clipboard, UI and input devices.
• Used by OpenClaw on Windows; NVIDIA OpenShell maps its agent policies to MXC-backed containment.
• Agent 365 native integration (preview in July) brings Defender, Entra, Intune and Purview to local agents.

Computer-using agents get secure, IT-managed Cloud PCs to execute multi-step enterprise workflows — opening apps, navigating interfaces, entering inputs and processing data — on a consumption model.

• Windows 365 Flex (formerly Frontline) reaches GA; new 32-vCPU and GPU-enabled "Select" SKUs.
• Windows 365 ships with the new Windows developer configuration (public preview).

A developer device powered by NVIDIA RTX Spark silicon delivering up to 1 petaflop of AI compute and 128 GB unified memory — enough to run up to 120B-parameter models locally with up to 1M-token context, without cloud GPU instances.

• WSL 2 with GPU passthrough and full CUDA support pre-configured, with VS Code and GitHub Copilot installed.
• Available later this year in the U.S. via Microsoft.com.

A wave of developer-experience upgrades make Windows more Linux-friendly and agent-aware out of the box.

• Coreutils for Windows (GA) — Linux-like CLI utilities built on the Rust uutils project, running natively.
• WSL containers (coming to preview) — build, run and manage Linux containers via a native CLI & API.
• Intelligent Terminal (experimental) brings context-aware agents into Windows Terminal via the Agent Communication Protocol (ACP).
• Aion 1.0 on-device SLMs — Instruct (everyday text intelligence, open weights) and Plan (14B reasoning/tool-calling, coming months).

A fully managed PostgreSQL service delivering ultra-low latency, rapid read scale-out, built-in AI and deep Azure integration — more than 3× the throughput of comparable self-managed setups in internal testing.

• Up to 192 vCores and 128 TB storage; positioned against Amazon Aurora and Google AlloyDB.
• Accelerated app development with GitHub Copilot in VS Code, in-database model access, vector indexing and semantic search.

A managed backend-as-a-service on Microsoft Fabric — database, auth and app hosting — defined through GitHub-based workflows so developers and coding agents move from prototype to production without managing infrastructure.

• A partnership with Replit creates a fast path from prototype to governed, enterprise-grade deployment.

GPU-powered acceleration for the Fabric Data Warehouse, reported as up to 27× faster than vCPU-based Fabric DW in Microsoft's testing.

Microsoft's first "autopilot" — an always-on personal agent for work, built on OpenClaw and Work IQ, that proactively handles meeting prep, scheduling conflicts and routine tasks without being asked.

• Operates across cloud, desktop and web, connecting to Teams, Outlook, OneDrive and SharePoint.
• Brought to Frontier customers; English-only, requires GitHub Copilot & M365 licensing, metered through GitHub tokens.

Programmatic access to the Work IQ intelligence layer (GA June 16), plus a Learning Agent that accelerates Copilot adoption with role-based, in-flow skilling included in M365 Copilot.

• Work IQ APIs are independent of individual user licensing and priced on consumption.

A single Microsoft Admin Center hub for IT to enable credit-based billing, set spend limits and monitor tenant-pooled Copilot usage across users, groups and agents.

An enterprise-grade agentic AI platform for the full scientific workflow — deploy autonomous agent teams, guided by human expertise, to reason over knowledge, generate hypotheses, optimise experiments and validate theories in a continuous loop.

• BHP is finding copper-leaching solutions in months instead of years; Syensqo is accelerating semiconductor R&D; GSK is iterating on drug discovery.
• A free Discovery local app (preview) runs on your computer with just a GitHub Copilot account.

Microsoft's next-generation topological quantum chip, developed with help from Microsoft Discovery's agentic AI: a 1,000× reliability improvement, with a mean qubit lifetime of 20 seconds (up to a minute).

• A new materials stack swaps aluminium for lead; one-microsecond operations and qubits ~1/100th of a millimetre.
• Microsoft now expects a scalable quantum computer by 2029 — half its original timeline.