Microsoft's Build 2026 Agent Control Plane: How the IQ Stack, Foundry, and Windows Create a Full-Stack Lock on Enterprise Agents

Microsoft used Build 2026 to declare that the company winning the AI agent market won’t be the one with the best model. It will be the one that controls where agents get their context, how they’re governed, and where they run. The June 2 keynote by Kyle Daigle, GitHub COO and Microsoft’s Chief Marketing Officer for Developer, laid out a platform thesis spanning silicon to cloud: agents built in GitHub, deployed to Microsoft Foundry, grounded in a four-layer intelligence stack called Microsoft IQ, running on Windows locally or Azure globally, with governance and audit trails embedded at every layer.

Ten days after the keynote, the architecture is worth examining in detail. What Microsoft announced is a control plane — the infrastructure layer that decides what an agent knows, what it can do, and who watches it do it.

The IQ Stack: Four Context Layers for One Agent

The centerpiece of Build 2026 is Microsoft IQ, a context architecture that gives agents access to four distinct knowledge sources through a unified retrieval system.

Work IQ captures organizational signals: emails, meetings, files, Teams messages, and the relationships between them. According to Microsoft’s official Build 2026 blog post, Work IQ APIs are scheduled for general availability on June 16, providing programmatic access to “how work actually happens across Microsoft 365, organizational systems and external sources: people, emails, documents, meetings and how they connect.”

Fabric IQ provides structured business data through Microsoft Fabric. Arun Ulag, Executive Vice President of Azure Data, described the role as solving the “context bottleneck” for enterprise agents: “Every new agent starts from zero, relearning how the business works, where data lives, and what rules to follow.” Fabric IQ connects agents to company ontologies, formal models of business entities and relationships linked to live data in OneLake.

Foundry IQ is the retrieval planner. Pablo Castro, CVP and Distinguished Engineer, announced that Foundry IQ knowledge bases are now generally available with “fully SLA-backed knowledge layer with stable APIs, compliance certifications, and the Foundry IQ MCP server for any MCP-compatible host.” The MCP integration matters: it means Foundry IQ can serve context to agents built on any framework that speaks Model Context Protocol, not just Microsoft’s own tooling.

Web IQ is the newest layer. Launched the same day, Web IQ is described as “a search engine for AI systems” built on the Bing index but re-architected for agentic workloads. It delivers grounding passages at “sub-165 ms latency” and claims 2.5x the speed of the next fastest alternative. The Bing blog post is explicit about the architectural shift: “Agents do not issue a single search and stop. They retrieve repeatedly, reason over evidence, adapt to new information, and operate inside tight latency budgets.”

Together, these four layers create a single retrieval surface that spans internal documents (Work IQ), structured business data (Fabric IQ), indexed enterprise knowledge (Foundry IQ), and the live web (Web IQ). An agent built on this stack can answer “What did our CFO say about Q3 margins in last Tuesday’s meeting?” and “What are current semiconductor tariff rates?” in the same reasoning chain.

Windows as Agent Runtime

The second Build 2026 pillar is less obvious but potentially more consequential: Windows as a local agent execution environment.

Daigle’s keynote framed this as “Windows for developers, period,” introducing a new developer configuration with local agent sandboxing, enhanced Windows Subsystem for Linux capabilities, and an “intelligent shell and terminal experience.” The WindowsForum analysis interpreted this as Microsoft positioning the OS itself as part of the agent control plane, providing local compute for agent tasks that don’t need cloud round-trips while maintaining the same governance model.

This is the pattern that distinguishes Microsoft’s approach from pure-cloud competitors. Google’s Gemini agents run in Google’s infrastructure. Anthropic’s Claude agents run through API calls to Anthropic’s servers. Microsoft is offering local execution on Windows machines with the same identity, permissions, and audit trail that govern cloud-deployed agents. For enterprises with data residency requirements or latency-sensitive workflows, this hybrid model addresses a constraint that API-only platforms cannot.

The Windows runtime also connects to Project Solara, Microsoft’s Android-based agent-first hardware platform announced at the same conference. Between Windows on desktops, Solara on purpose-built devices, and Azure in the cloud, Microsoft is building a three-surface runtime grid.

Scout: The Platform Thesis in Product Form

Microsoft Scout, introduced by Corporate Vice President Omar Shahine, is the first product that demonstrates the full control plane in action. Scout is an always-on personal agent that handles meeting prep, scheduling conflicts, and routine tasks proactively.

Two architectural decisions stand out. First, Scout runs on OpenClaw, making it one of the highest-profile enterprise deployments of the open-source agent framework. Microsoft chose an open runtime rather than building a proprietary execution layer. Second, Scout is grounded in Work IQ, meaning it has access to the full organizational context layer: emails, documents, calendar events, and team relationships.

The combination is instructive. The agent runtime is open (OpenClaw). The context layer is proprietary (Work IQ). The governance layer is proprietary (Foundry). This is how platform lock-in works in the agent era: you can leave the runtime, but your agents lose their memory and context when you do.

The Model Layer: MAI-Thinking-1 and Model Diversity

Microsoft’s model strategy reinforces the platform thesis. The company released MAI-Thinking-1, a 35 billion active parameter reasoning model with a 256K context window, trained on “enterprise grade, clean and commercially licensed data.” According to the Build keynote blog, independent raters preferred it to Sonnet 4.6, and it matches Opus 4.6 on SWE Bench Pro coding benchmarks.

But the more significant announcement was model diversity as explicit strategy. Fireworks AI reached general availability on Foundry, and MAI models will also be available through Baseten and Open Router. Daigle’s framing was specific: a platform “that’s model diverse, open and heterogeneous at every layer of the stack.”

This is Microsoft deliberately avoiding the single-model bet. If the platform owns context, governance, and runtime, it doesn’t need to own the model. Customers can use MAI-Thinking-1, Claude, GPT-5, Gemini, or any open-source model. As long as they deploy through Foundry and ground through the IQ stack, Microsoft captures the infrastructure revenue regardless of which model runs inference.

Meanwhile, Microsoft also announced Skill Opt, a framework linked to the AutoGen multi-agent orchestration platform that targets token waste in multi-step agent tasks. As TechnoSports reported, the framework “aims to eliminate redundant function calls” by filtering which skills an agent actually needs at runtime. The token economics angle is critical: if agents cost enterprises $10,000/month in token consumption, a framework that cuts that cost by 30-50% becomes a compelling reason to stay on the Microsoft platform.

The Consolidation Trade-off

The strategic risk Microsoft is asking enterprises to accept is consolidation at a depth the industry hasn’t seen since the Windows Server and Active Directory era of the 2000s.

An enterprise that adopts the full stack gets agents grounded in their organizational knowledge (Work IQ), their business data (Fabric IQ), indexed knowledge bases (Foundry IQ), and the live web (Web IQ). Those agents run on Windows locally, Azure globally, and Solara on purpose-built devices. They’re governed through Foundry, audited through Azure, and deployed through GitHub.

The value proposition is real: no other vendor offers this breadth. Google has strong models and cloud infrastructure but no enterprise OS or productivity suite with comparable market share. Anthropic has strong models and API governance but no data platform, no OS runtime, and no enterprise knowledge graph. OpenAI has consumer reach and model capability but has shown limited interest in the governance and context layers that enterprise IT departments require.

But the lock-in cost is equally real. As the WindowsForum analysis noted, “a decision efficient in 2026 could look restrictive in 2028 if model capabilities or pricing shift.” Work IQ’s value derives from integration with Microsoft 365. Fabric IQ’s value derives from data already in OneLake. Foundry IQ’s MCP support provides some portability, but migrating a production knowledge base with organizational context, permission models, and compliance certifications is not a weekend project.

The Platform Race in June 2026

Microsoft’s Build 2026 strategy reveals a market that has split into three distinct competitive layers.

The model layer is commoditizing. MAI-Thinking-1 matches Opus on coding benchmarks. GPT-5 and Gemini 3 offer comparable reasoning. Open-source models from Meta and Mistral close the gap on the low end. Price cuts are accelerating across all providers.

The runtime layer is fragmenting. OpenClaw, LangChain, CrewAI, AutoGen, and dozens of smaller frameworks compete for developer adoption. Microsoft’s bet on OpenClaw for Scout suggests even Microsoft doesn’t think proprietary runtimes are the defensible layer.

The context and governance layer is where the winner-take-most dynamics concentrate. Whoever owns the organizational knowledge graph, the permission model, the audit trail, and the compliance certifications owns the switching cost. Microsoft’s IQ stack is the most complete attempt to own this layer.

The question for enterprise buyers is whether they want one vendor controlling context, governance, runtime, and OS for their agent infrastructure, or whether the risks of fragmentation (higher integration cost, less seamless context) outweigh the risks of consolidation (vendor dependency, pricing leverage, reduced optionality).

Ten days after Build 2026, no competitor has announced a comparable full-stack agent platform. That gap is either Microsoft’s window of opportunity or the industry’s warning signal, depending on which side of the procurement table you sit on.