Spoilers …

Microsoft spent Ignite talking about something larger than chatbots. Copilot as a work surface, Microsoft IQ as a context layer, Agent 365 as a governance plane, Perception as a security response layer, and Azure/Foundry as the execution platform. The architectural direction is compelling. The operational question is harder. Can we inventory the agents, control their identities, explain their costs, prove their actions, and recover when they go wrong?

My take: the headline is not “AI replaces IT.” It is that IT inherits an entirely new estate to operate. One full of non-human identities, ephemeral execution, dynamic permissions, data boundaries, inference bills, and autonomous change. If we skip the fundamentals, we will automate our existing operational debt at machine speed.

Announcements I would actually pay attention to

Copilot stops being a sidecar

Home (Chat + Cowork) and Office in Copilot are GA as of November 17. Rather than merely popping open Word, Excel, or PowerPoint beside a chat, Microsoft is placing native document creation and collaboration inside the Copilot experience. Copilot Managed Runtime is also GA; Copilot Code remains in Frontier preview with December GA targeted, while Autopilot and Copilot Workspaces are Frontier experiences.

The question: How do the resulting apps and agents get identities, owners, lifecycle policy, data classification, audit trails, and a support model? A beautiful demo is not a service catalog.

Copilotforce: Practical cross-cloud integration, not just a shiny connector

Microsoft and Salesforce announced Copilotforce: Salesforce in Copilot (public preview, 38+ sales skills), Sales Cloud in Teams (pilot), and Salesforce Data 360 / Fabric Zero Copy integration. Important nuance: Microsoft identifies the Data Federation capability for supported OneLake data as GA; that does not mean every aspect of the broader two-way integration is GA.

For architects, this is a permissions-and-data-lineage project. Test whether Salesforce permission boundaries actually survive MCP calls, and identify which system is authoritative for sales, finance, and customer attributes.

Microsoft IQ: “your IQ” is the real strategic play

Microsoft IQ is GA as an umbrella platform, connecting Work IQ, Fabric IQ, Foundry IQ, and Web IQ. But each component has separate readiness: SharePoint indexing for Foundry IQ is GA; knowledge graphs and Microsoft-Retriever-1 are preview; Fabric semantic-model usage in Copilot is GA, while selected Excel and GitHub experiences are preview. Business plugins under the Frontier program add reusable, hierarchical instructions and policy.

My take: this is knowledge architecture, not merely prompt engineering. It raises familiar engineering issues: source-of-truth design, entitlements, stale documents, provenance, quality gates, and change control. This is where the value will either compound or collapse.

Models: MAI, GPT 6.2, and the economics of routing

GPT-6.2 was announced for Microsoft Foundry alongside expanded Meta model availability. Microsoft-Decision-1.1 is GA. MAI Code 2 and Code 2 Flash are GA within GitHub Copilot; Microsoft separately describes MAI Code 2 Flash as limited-access preview in Foundry. MAI Cyber 2 Flash, MAI Image 2.7, and Image 2.7 Flash add purpose-specific options. MAI Licensed Weights offers qualifying Frontier customers greater customization control.

Microsoft reports Code 2 reaching about 98% of GPT-6.1 Sol coding benchmark performance at 40% of the price. That is a vendor-reported benchmark, not a transferable guarantee. Use your own tasks, quality metrics, latency measurements, and total cost per successful result before updating production routes.

Agent 365 + Perception: non-human identity meets non-human action

Microsoft reports more than 50 million registered enterprise agents. Agent 365 supplies inventory, governance, and visibility, while Business Analytics begins in Frontier preview for cost, adoption, and assisted-value analysis. Perception is GA, with red/blue/green security agents and playbooks. Several new Perception features, including AI-agent risk playbooks and web-app testing, have narrower preview gates.

For security teams: an agent is an insider with software speed. It needs explicit ownership, least privilege, monitored tool calls, revocation, runtime containment, reproducible traces, and human approval for consequential actions. The AI Transparency Initiative is a standards effort, not a new GA enforcement switch.

Release-readiness matrix as of November 17, 2026

Status means the stated state of the specific capability, not an entire product brand. “Today” denotes announcement/availability on Nov 17 only where Microsoft states it; preview enrollment, tenant rollout, region, and licensing may vary.

Azure and hybrid infrastructure: the quieter, bigger story

Silicon and performance are moving toward the agent stack

Cobalt 200 Arm-based VMs are GA in 11 named Azure regions, with 24 regions targeted by the end of 2026. Microsoft claims up to 50% more CPU performance than Cobalt 100. Maia 300 remains a 2027 roadmap item rather than a generally available product. Boost DPU, HSM, networking, and model execution should be evaluated as a complete unit of cost, not isolated hardware marketing claims.

Azure HorizonDB is GA; the Fabric story has a separate release gate

Azure HorizonDB brings managed PostgreSQL compatibility, multi-zone resilience, and built-in vector capabilities to GA. Fabric HorizonDB, the serverless PostgreSQL experience inside Fabric, is preview; database mirroring from Azure HorizonDB to Fabric is also preview. Keep that distinction in architecture decision records and production timelines.

AKS is becoming the governed execution boundary

Agent Grid, the AKS Agent Sandbox extension, and Foundry-to-AKS model/agent deployment workflows were presented as previews. The pattern is attractive: let Foundry support development and evaluation while AKS provides organizational placement, runtime boundaries, networking, and controls. The challenge is no different from standard platform engineering: how are images signed, secret access minimized, runtime policies tested, logs retained, and node pools costed?

Confidential, sovereign and local: workload placement now matters more

The new Azure confidential enclave, Kata CoCo, and multi-GPU confidential compute capabilities are in preview, while AMD v6 confidential compute for Azure Database for PostgreSQL is GA. Windows Server 2028 is in preview, and its hybrid AI capabilities have a limited early access path. Microsoft’s sovereign-cloud direction reinforces a placement decision: data residency, model locality, attestation, and control boundaries need to be stated per workload.

My five operational takeaways

Treat agents as managed identities and assets.

Create an agent CMDB/inventory: owner, business purpose, workload identity, tool permissions, data classification, allowed environments, expiration date, incident runbook, and decommission procedure.

Put a policy gate between intent and execution.

Require approvals for destructive operations, external sharing, privileged changes, financial actions, and production remediation. Capture tool inputs, outputs, and decisions as evidence.

Build a FinOps for AI baseline before scaling.

Track tokens, model routing, per-task success, agent runtime, hosted compute, observability, storage and rework. A cheap token is expensive when it creates an incorrect result.

Keep the architecture portable.

Publish source in Git, store IaC and policy alongside the workload, test model substitutions, use portable protocols where viable and document where Foundry, M365 and Salesforce dependencies create lock-in.

Use preview strategically, not accidentally.

Separate experimentation subscriptions/tenants from production; use explicit change records for preview onboarding, data retention, feature flags, support commitments, and exit criteria.

A practical 30-60-90-day Ignite response

Small announcements worth watching

GitHub Copilot SDK 2.0 is GA and may matter more to teams building durable agent services than a flashy demo. Azure Monitor health models are GA, providing architecture-aware incident context; Azure SRE Agent upgrades are in preview. Azure Chaos Studio Workspaces and its AKS extension are GA; use them to test the resilience assumptions that autonomous operations will otherwise hide.

The /cost command in Copilot, Microsoft Today, and Copilot Fusion are worth investigating, but avoid treating conference demonstrations as a guarantee of tenant-wide entitlement. Verify the product/feature gate and licensing in the admin center before making a business commitment. Official Book of News confirms Today as a private preview and distinguishes Frontier Copilot experiences from GA products.

Final thoughts: The control plane is the product

Ignite 2026 did not convince me that every organization is suddenly ready to become a frontier company. It convinced me that Microsoft is assembling an unusually broad enterprise AI stack. That makes the architecture interesting, and makes operational discipline non-negotiable.

The organizations that benefit will not be the ones with the most agents. They will be the ones that can tell you who owns every agent, exactly what it can touch, what it costs, what it changed, and what evidence proves it did the right thing.

If your agent ran a production change tonight, would you have the evidence by tomorrow morning?