Microsoft opened Ignite 2026 in Seattle this week with a message that had surprisingly little to do with model releases. The frontier models, the company argued, are commoditised. The real fight is over the layer underneath them: the operating system for enterprise agents, the context they need to be useful, and the governance that lets organisations trust them. Across Azure, Foundry, Fabric and a new platform called Microsoft IQ, the announcements form the most substantial enterprise agent infrastructure release since the category began to take shape.

The shift from model-first to operations-first thinking is the defining story of enterprise agentic AI in late 2026. Microsoft's Ignite keynote did not lead with a new large language model. It led with an observation that Satya Nadella first made at Build in June: the company added more datacentre capacity in 18 months than in Azure's first decade combined. The message was that infrastructure, not intelligence, is now the binding constraint on enterprise agent deployment.

The agentic cloud: from read-only to full control

The most consequential infrastructure announcement is the Azure Resource Manager MCP server. Previously read-only across more than 3,500 resource types, it can now create, update and delete resources, with every request gated by Azure Policy. For enterprises deploying agents that need to provision infrastructure, scale workloads or respond to incidents, the change is fundamental. An agent that can only observe its environment is a dashboard. An agent that can act on it is operations.

Alongside it, Azure Copilot's deployment agent reached general availability, accepting natural-language descriptions and delivering deployed, governed infrastructure without a human writing a single configuration file. The Azure SRE Agent picked up enterprise controls in preview, including customer-managed encryption keys, Microsoft Entra Agent ID integration and cross-agent communication. Together, the three capabilities form what Microsoft is calling agentic cloud operations: agents that do not just advise on infrastructure but manage it, within policy boundaries that the organisation controls.

Decision-1.1: a model built for agent routing, not conversation

One of the more technically interesting announcements is Decision-1.1, now generally available. It is a fast, cost-effective model designed to make decisions in software at scale, rather than conversing with a user or generating content. Where large language models spend tokens thinking through routing, classification, prioritisation and workflow control, Decision-1.1 reaches answers within constrained definitions and provides calibrated confidence scores, letting applications defer to a human when uncertainty is high.

The practical implication for enterprise buyers is significant. Every multi-agent system contains a decision layer that routes tasks, verifies outputs and controls workflow. Running that layer through a general-purpose model is expensive and introduces variability. Decision-1.1 is built to handle that layer at lower cost and with more predictable behaviour. It is built on Google's Gemma 4 26B A4 and adds multimodal support, so it can analyse images in instructions alongside text.

Microsoft IQ: context as enterprise intellectual property

The second pillar of the Ignite message is context. Microsoft IQ bundles Work IQ, Fabric IQ, Foundry IQ and Web IQ with more than 350 connectors into one enterprise context management platform. The premise is straightforward: an agent that starts every task from zero, with no knowledge of the business it serves, is limited to generic outputs. An agent that can pull contract terms from Fabric IQ, identify affected orders, find the responsible owner through Work IQ and check the relevant policy in Foundry IQ is performing enterprise work.

The supply-chain example Microsoft demonstrated shows the intent. One business plugin identifies affected orders, checks contract terms, and finds the responsible owner, all without a human navigating between systems. The context stays within the organisation's Azure tenancy. Microsoft's framing is that context is intellectual property, and enterprises should treat it the same way they treat their customer data, their financial records and their source code.

Foundry and the continuous improvement loop

Microsoft Foundry now supports more than 11,000 models from OpenAI, Anthropic, Meta, Mistral and the open-source community, with a model router and AI Gateway for intelligent model selection. But the more important update is the addition of a continuous improvement workflow. Internal evaluations let autonomous agents establish baselines for quality, safety, latency and cost. Human annotations, now in preview, let domain experts rate and comment on agent traces based on business judgement, feeding back into an improvement loop. Optimisation tests compare agent behaviour against evaluations and real-world results, and each improvement becomes the next baseline.

For enterprise procurement teams, this matters because it answers a question that has been quietly eroding confidence in agentic AI: how does an agent get better over time, and who is responsible for making that happen? The Foundry loop provides a structured answer, with tooling that makes improvement observable and measurable rather than assumed.

The model portfolio: purpose-built, not just general-purpose

Microsoft also extended its first-party MAI model family, with clear segmentation by task and cost. MAI-Code-2 handles demanding agentic coding work, while MAI-Code-2-Flash supports high-volume, everyday coding at lower cost, letting development teams choose between performance and economics. MAI-Image-2.7 and MAI-Image-2.7-Flash offer sharper editing and more consistent image generation for brand teams managing visual output at scale. MAI-Cyber-2-Flash is a purpose-built model for security scanning, available in preview inside Codename MDASH, Microsoft's agentic code security system.

The portfolio approach is itself a signal. A general-purpose model that tries to excel at code, images, security and decisions is a compromise in every direction. Purpose-built models that handle one domain well, at a predictable cost, are a more reliable foundation for enterprise-grade agent systems.

Hardware that matches the workload

Azure Cobalt 200 virtual machines reached general availability, delivering up to 50 percent better CPU performance than Cobalt 100 and targeting scale-out, cloud-native and agentic AI workloads. Microsoft pairs its own Maia and Cobalt chips with silicon from NVIDIA and AMD, placing workloads based on available capacity so GPU investment stretches further as demand grows. The implication for enterprises running agent workloads at scale is that infrastructure cost is becoming a tunable parameter rather than a fixed overhead.

What it means for enterprise buyers

For procurement, architecture and operations teams evaluating agentic AI platforms, the Ignite announcements carry four practical messages.

First, the operating layer is becoming a product category. The combination of writable MCP infrastructure, deployment agents that translate natural language into governed environments, and SRE agents with enterprise controls means that operating AI agents is no longer a manual, bespoke activity. It is a platform capability with defined boundaries, audit trails and policy enforcement.

Second, context management is now infrastructure. Microsoft IQ treats enterprise knowledge as a governed asset that agents can access through connectors, not as a training corpus to absorb. For organisations in regulated industries, this distinction matters. An agent that queries the source of truth in real time is safer and more auditable than one that has memorised it during training.

Third, the decision layer deserves its own model. Decision-1.1 signals that agent routing, classification and workflow control are workloads that benefit from specialised, predictable models rather than general-purpose inference. Buyers building multi-agent systems should ask whether their decision logic runs on the right infrastructure or whether they are paying general-purpose prices for routing work.

Fourth, continuous improvement is now part of the platform, not a separate project. The Foundry evaluation-and-annotation loop gives organisations a structured way to measure, improve and document agent performance over time. For buyers under pressure to demonstrate ROI, that is an operational necessity, not a nice-to-have.

What it means for suppliers

For companies selling agentic AI platforms, frameworks or services, Ignite 2026 resets expectations in two important ways.

The first is that the largest enterprise cloud platform is now shipping the operating layer as part of the product, not leaving it as an exercise for the customer. Specialist agent platforms that do not offer equivalent governance, context management and continuous improvement tooling will face harder questions in enterprise procurement cycles where Azure is the incumbent infrastructure choice.

The second is that purpose-built models are becoming a differentiator. A platform that routes every decision through a general-purpose model, without offering a lower-cost, higher-reliability alternative, is architecturally behind one that segments its workload by model type. Suppliers should expect buyers to ask which model handles which task, and at what cost, with evidence rather than estimates.

The Agentic Expo angle

Agentic Expo exists because the enterprise AI market has moved past the question of whether agents work. The question now, as Ignite 2026 demonstrated, is whether enterprises can operate them at scale, govern them with confidence, and measure their contribution to business outcomes.

At Olympia London on 23 to 24 March 2027, the suppliers building the operating layer for enterprise agents will meet the buyers who need to deploy it. The conversations will not be about model benchmarks. They will be about policy enforcement, context management, cost control and the operating model that turns an agent from a pilot project into a production asset. That is the work Microsoft Ignite is now shipping as product. It is also the work Agentic Expo is built to advance.

Get Tickets Exhibit at Agentic Expo

Sources: Microsoft, "Microsoft Ignite 2026: What's new in Azure, the platform purpose-built for AI", 18 November 2026; SiliconANGLE, "Microsoft rolls out new first-party models and updates to agentic AI continuous improvement", 17 November 2026.