Microsoft Build 2026: What IT Leaders Need to Know
Microsoft Build took place from June 2nd to June 3rd in San Francisco. The central theme was agents moving from demonstrations to production tools. As we mentioned in our recent Big 5 AI Vendor Roundup, Microsoft presented seven new MAI models “across image, voice, transcription, coding, and reasoning.” But Microsoft also positioned itself as providing an integrated, governed agent platform. The strategy is to emphasize the system around AI rather than individual models.
In the opening keynote, CEO Satya Nadella asked, “How do you all participate fully in this frontier intelligence ecosystem?” The answer lay in Microsoft’s announcements, which covered the entire agent stack, including data, build-and-run, governance, billing, and devices.
Mapping the Platform
Microsoft presented an agent platform strategy built around connected products rather than isolated tools. The message was clear that Microsoft wants companies to adopt one governed agent system instead of throwing together a bunch of separate tools.
The vision for the Microsoft Agent Platform is that GitHub is where developers build, so the GitHub Copilot app, now in preview, will enable developers to manage agentic development workflows on their desktop. As part of the governed agent lifecycle, Microsoft IQ supplies business and work context to agents. Agents then run in Foundry with access to the majority of coding models and are governed by Agent 365, which is essentially an IT admin console for agents.
At Build, Microsoft positioned Foundry as a central platform for building and operating production agents, not just a model catalog. Agents can then reach users through Microsoft 365 Copilot, Teams, and other Microsoft apps.
The Personal Productivity Agent Lands on the Desktop
OpenClaw is an open-source, self-hosted framework for autonomous agents that can execute multistep workflows, control apps, browse the web, and manage files. We covered its release in a note from earlier this year. At Build 2026, with OpenClaw creator Peter Steinberger on stage, Microsoft announced that it is integrating OpenClaw into its agent stack through a more controlled Windows runtime environment. It runs on Windows leveraging Microsoft Execution Containers (MXC), a policy-driven runtime layer that restricts what agents can access. On stage, this was demonstrated through the agent not being able to delete all files on a desktop even when being instructed to do so by the user. Steinberger enthusiastically stated: “You can run OpenClaw inside your company now.” In a curious turn of events, OpenClaw has developed from being an enterprise no-go into a secure, sandboxed tool. For enterprise users, Microsoft Scout is Microsoft’s first “Autopilot,” an always-on personal work agent built on OpenClaw and Work IQ. It works across Microsoft 365 apps such as Teams, Outlook, OneDrive, and SharePoint. There is also a desktop app for Windows and macOS. Currently, Scout is only available through the Microsoft Frontier program. Access is controlled by IT through Frontier enrollment, Intune policy configuration, admin attestation, and applicable Copilot licensing. Access is controlled by IT through Frontier enrollment, Intune policy configuration, admin attestation, and applicable Copilot licensing.
Earlier this year, an OpenClaw agent ran out of control inside a researcher’s inbox, illustrating the risk of highly autonomous agents acting across mail or files without sufficient controls. Scout’s enterprise controls and MXC-style containment are intended to address that risk, but by no means do they remove the need for governance.
One thing the announcement did not resolve is cost. The attraction of Autopilot agents is that they can run in the background and take action when needed, but that is also what could make them expensive. A persistent agent that performs multi-step work over long periods consumes far more tokens than a simple chatbot interaction.
Foundry Roadmap: Measurement Is Yet to Come
Microsoft Foundry is becoming the main environment to build and run enterprise agents. But the tools to build and run agents are becoming available first. The tools to control, measure, and optimize them are still maturing through 2026.
First, hosted agents in Foundry Agent Service are expected late June to early July, with sandboxed sessions, persistent state, filesystem access, and enough framework flexibility that you're not locked into one toolkit. The Foundry Toolkit for VS Code is already GA, as are Foundry IQ knowledge bases, which provide an SLA-backed retrieval layer.
The rest is a preview or a promise. Publishing agents into Teams and Microsoft 365 Copilot is planned for GA this June. Toolboxes (one governed front door for tools, skills, MCP clients) and enterprise data, are in public preview, and so are scheduled routines and agent memory.
Where it gets a little thinner, or muddier, is the tools that tell you whether an agent actually behaved. Guided Guardrail Setup, the Rubric quality evaluator, and framework-agnostic tracing are all in preview. Agent Optimizer is "coming soon." Agent ROI, the one your CFO will eventually ask about, is in private preview. The two open specifications Microsoft put on the table are Agent Control Specification, which drops fixed checks into an agent's run, and ASSERT, which turns your written policies into tests an agent must pass. But new specs only matter if people adopt them.
So, before you build any of this into a 2026 production plan, check whether the piece you're counting on is actually GA or a promise.
Nothing New (Yet) on the Question of Cost
Billing was not the headline at Build 2026, but it is embedded in the agentic strategy. Microsoft’s agent usage is increasingly governed by Copilot Credits, which is a metered currency used across Copilot Studio agents, Dynamics 365 agents, Microsoft 365 Copilot Chat sessions for unlicensed users, and Work IQ APIs.
Copilot Credits is not new, since Microsoft changed the agent currency from messages to Copilot Credits on September 1, 2025. Credits are obtained through pay-as-you-go meters, prepurchase plans, or prepaid packs. The credits report measures usage from people without a Microsoft 365 Copilot license who interact with agents.
At the moment, Microsoft 365 Copilot licensed users can use many employee-facing Copilot Studio agents in Teams, SharePoint, and Copilot Chat without extra Copilot Credit charges. But given Microsoft’s broader move toward usage-based pricing, including its recent changes to GitHub Copilot, organizations should start tracking what internal agents use, not only external or customer-facing agents. The real risk lies in multiplication, because while a chatbot may burn a small amount, what about an always-on agent that’s running multistep work for hours?
Our Take
Build 2026 was primarily about productizing and governing agents, not about model capability. The components to build and run agents are becoming available now, while the components to govern, measure, and optimize them are mostly still maturing. As Microsoft moves toward usage-based pricing, the cost of these agents is also something worth watching.
- Plan against the announced direction for agent governance components and confirm timing and licensing for each.
- Inventory who has access to Copilot or Scout agents and what data they can reach.
- Before adoption, decide an organizational position on always-on agents touching mail and calendars.
- Model Copilot Credits as a usage cost line and set internal guidance or policies for agent selection before granting access.
Want to Know More?