“Agents are not just a feature. They are the new operating system for work.” Satya Nadella delivered that line 12 minutes into the Microsoft Build 2026 keynote on June 2 at Fort Mason Center in San Francisco, and the next two hours made it concrete. Office 365 Copilot now defaults to Agent Mode. GitHub Copilot gained an autonomous coding agent. Windows ships on-device AI models on June 9. And Project Polaris, Microsoft’s in-house coding model, replaces GPT-4 Turbo in Copilot starting August 2026.
Five thousand developers and IT professionals attended in person. Four hundred technical sessions followed the keynote. Every major announcement centered on one concept: agents that plan, act, and complete tasks across applications without waiting for a prompt.
Office 365 Copilot Agent Mode Is Now the Default
The centerpiece announcement is also the most immediately consequential for enterprise users. Copilot Agent Mode is now the default interface in Word, Excel, and PowerPoint, rolling out to all Copilot for Microsoft 365 subscribers by late June 2026.
Agent Mode shifts Copilot from a synchronous assistant (you ask, it answers) to an asynchronous coworker that executes long-running tasks across the entire suite. Jared Spataro demonstrated three simultaneous agents during the keynote: a Contract Review Agent parsing legal documents in Word, an Excel Forecasting Agent running scenario analyses, and a Scheduling Agent coordinating calendars in Outlook. Each agent maintains its own context, permissions, and memory. Agents show up as named participants in Teams channels.
Memory persistence is governed by Microsoft Purview data loss prevention policies. A custom agent marketplace opens in Q4 2026. The technical stack runs on GPT-5.5-turbo for cloud inference and Phi-4-mini for on-device tasks, with a target response time under 300 milliseconds for simple agents.
Microsoft reports that 67% of Fortune 500 companies already use at least one Copilot feature, and agent capabilities have been the single most requested enterprise feature for 18 months. For the 400 million commercial Office users, this changes the daily interaction model. Copilot is no longer a sidebar you invoke. It is a set of persistent agents running in the background, completing tasks you delegate.
GitHub Copilot Goes Autonomous in July
GitHub Copilot Enterprise subscribers gain Autonomous Agent Mode starting July 2026. The evolution from autocomplete to full coding agent is now complete.
In a live demo, a developer typed a single natural language prompt: “Build a .NET MAUI cross-platform settings page with dark mode toggle and MQTT telemetry.” Within 10 minutes, the agent generated a 12-file pull request with unit tests that passed on the first run.
Every autonomous change still requires human approval before merging. A new Agent Sandbox spins up an ephemeral Linux container for each task, isolating the agent from production repositories. A GitHub Compliance Scanner checks generated code against 50+ security and licensing policies before any pull request opens.
Pricing stays on the existing $39/user/month Enterprise tier, with consumption measured in Copilot Compute Units (CCUs). Each feature request burns an estimated 50 to 75 CCUs, and teams receive 2,000 free CCUs monthly.
This puts GitHub Copilot in direct competition with Cognition’s Devin, which raised at a $26 billion valuation just last week. The difference: Copilot Autonomous lives inside the GitHub ecosystem where 100 million developers already work. Copilot Workspace also graduated from beta to general availability at Build, adding fleet mode and autopilot mode for autonomous multi-repo operations.
Project Polaris Replaces GPT-4 Turbo in Copilot
The quieter announcement with the biggest strategic implications: Project Polaris, Microsoft’s in-house coding model, replaces GPT-4 Turbo as GitHub Copilot’s default starting August 2026.
Polaris uses a mixture-of-experts architecture with specialized sub-modules for different programming languages and frameworks. It runs on Microsoft’s custom Maia AI accelerators in Azure, cutting per-inference latency and lowering inference costs compared to the OpenAI arrangement. According to Microsoft, Polaris outperforms GPT-4 Turbo on both HumanEval and MBPP benchmarks, with particular strength in low-resource languages like Rust and Haskell.
The migration is automatic for all subscribers, with a three-month fallback window for teams that prefer GPT-4. Pro tier subscribers gain multi-file context up to 100,000 lines and autonomous test generation.
The strategic read is unmistakable. Microsoft has been building its own MAI model family since April 2026. Project Polaris extends that independence to the company’s most valuable developer product. Microsoft now controls the model, the inference infrastructure, and the developer experience end to end. OpenAI’s own multicloud pivot made this separation mutual. The $13 billion partnership that launched the AI era is unwinding into competition, and Build 2026 put the sharpest date on the split: August, when GPT-4 loses its default seat in the world’s largest developer tool.
Windows Local AI: 1.4 Billion Devices Become Agent Hosts
Windows Local AI ships June 9 via Windows 11 KB5039239. It transforms Copilot+ PCs with compatible NPUs into on-device agent platforms.
The system runs models entirely on neural processing units from Snapdragon X Elite, Intel Lunar Lake, and AMD XDNA chips, requiring a minimum of 40 TOPS processing power. Three bundled models ship out of the box: Phi-4-mini-silicon (2B parameters) for text tasks, Phi-4-vision-silicon (7B) for image understanding, and a 300M-parameter speech-to-text model. No internet connection required.
Agents run inside a Windows Local Sandbox with no network access by default. File and microphone permissions require explicit user authorization through the Windows Copilot Runtime API. During the keynote, Microsoft demoed a local Meeting Recap Agent that analyzed a full Teams transcript in under two seconds without any data leaving the device.
The Windows Agent Framework 1.0 shipped with an MIT license, making agents first-class citizens of the OS. A Windows Agent Store launches alongside it with an 85% developer revenue share. Early partners include Adobe (Photoshop plugin) and Cisco (Webex on-device background blur).
Microsoft’s installed base of 1.4 billion Windows devices is the distribution advantage no other agent platform can match. Apple has the hardware integration. Google has the search traffic. Microsoft has the enterprise desktop, and now every qualifying PC is an agent host.
Azure AI Foundry Gets Agent Orchestration and Confidence Scores
Azure AI Foundry gained an Agent Orchestrator in preview (August 2026) that load-balances across thousands of agents. It supports heterogeneous agent teams built with Semantic Kernel, LangChain, or plain REST APIs and integrates with Azure Logic Apps for workflow automation.
The governance feature to watch: Agent Confidence Scores. The system assigns a percentage reliability rating to each agent’s output based on historical accuracy. Agents scoring below 95% automatically route their output to a human reviewer before actions execute. Every agent action carries a cryptographic attestation signature enabling retroactive policy verification.
Azure Container Apps offers a new Serverless Agent Hosting plan at $0.05 per agent-hour, with the first 10 agents free. For enterprise IT teams still evaluating what AI agents actually are and how to govern them at scale, this is the most practical announcement from the entire conference.
What Build 2026 Tells Us About the Agent Race
Microsoft made five major product announcements on one stage, and every one centered on agents. Not chatbots. Not assistants. Agents: software that plans, decides, and acts across applications.
The pattern matches what the entire industry has converged on this year. OpenAI’s GPT-5.5 pivoted from chatbot to agent platform. Google’s Gemini Intelligence turned Android into an agent OS for 3 billion users. Anthropic’s Dreaming gave agents the ability to learn from their own mistakes. Now Microsoft makes the same bet, but with a distribution channel (Office + Windows + GitHub + Azure) that touches more enterprise workflows than any competitor.
The vibe coding movement gets a corporate upgrade when the same AI that generates pull requests on GitHub also runs locally on a developer’s laptop with zero cloud dependency. The AI coding tool landscape just consolidated: Copilot now plays in the autonomous tier alongside Devin and Claude Code, not just the autocomplete tier.
Build 2026 was not a features conference. It was a declaration that Microsoft’s entire product surface, from Excel to VS Code to Azure to Windows itself, now runs on agents. Whether enterprises are ready for that is a different question. The platform already moved.