Between 17 and 20 November, Microsoft shipped agent announcements out of at least seven separate product groups: Entra, Agent 365, Foundry, Defender, Purview, Azure networking and Azure SRE Agent. Each group published its own blog with its own status labels. None of them published the architecture.
The stack does have a shape. Identity at the bottom, a management plane above it, a build and run layer, an isolation layer under that, detection alongside, operations on top, data governance across all of it and an open-source framework underneath. Most layers now hold something you could put into production this quarter. The preview labels cluster at the connections, and connections are what an adoption decision rests on. For what to act on first rather than how it fits together, we published a separate act, plan, watch and ignore triage. Where two Microsoft pages disagree below, both readings appear.
Identity: Entra Agent ID
This is the only layer with no price attached. Learn states that Agent ID is available for all Microsoft Entra customers. It gives you agent identity blueprints, agent identities with parent-child relationships and agents' user accounts, over OAuth 2.0, MCP and agent-to-agent. Automated lifecycle management of those identities is generally available, with an out-of-box policy that disables agent identities inactive for 60 days.
Nearly everything built on top is preview. Agent ID for discovered local agents, Agent ID for digital teammates, Conditional Access for assistive agents, dynamic permission remediation and access packages for digital teammates are each preview per the Entra Ignite blog. Conditional Access support for Windows 365 for Agents is generally available. Our deep dive on what Entra Agent ID records for a coding agent found the sign-in and audit trail real and usable while the controls around it stay prerelease. It is evidence today, not enforcement.
Two limits matter for the layers above. Learn says extending Microsoft Entra security features to agents requires Microsoft Agent 365, included with Microsoft 365 E7 and sold as an add-on to E5, A5 or Business Premium. And the documented Microsoft product integrations that create identities automatically are Copilot Studio, Security Copilot and Azure AI Foundry. Everything else arrives through Graph, CLI, infrastructure as code or a user consent prompt, and blueprint principals are capped at 250 agent identity creations each.
Management: Agent 365
Agent 365 is generally available for the Commercial segment on a per-user basis, and the Learn overview links that status to an announcement dated 1 May 2026. Microsoft E5 is the recommended prerequisite, and at least one user must hold a qualifying licence to enable it. The agent registry in the Microsoft 365 admin center is the inventory, and the tools registry and governance of plugins and skills are generally available.
The preview list is specific. Multi-tenant agent management is public preview, agent map use cases and local and shadow agent governance are preview, the Action Center arrives in preview in December, and external catalogs in the tools registry reach private preview in December.
What Agent 365 does not do deserves more attention than it got. Connected platforms covers Amazon Bedrock, Google Vertex AI, Salesforce Agentforce, Databricks Genie, Anthropic Claude Managed Agents, Oracle Generative AI Agents and Snowflake Cortex, with Oracle and Snowflake observability marked "Coming soon". Learn lists three limitations on the read-only connection. Synchronisation is manual, with no schedule. Deletion reconciliation is not guaranteed, so an agent removed from the source platform may stay listed. And registry sync adds no code-level capability, for which you need the Agent 365 SDK. Observability data is kept 30 days, and Agent 365 is not yet a Core Online Service under the Microsoft Product Terms.
Build and run: Foundry
The developer path is the most production-ready part of the stack. The Foundry Dev Pack went generally available on 18 November and bundles the Foundry CLI, generally available since 5 November, with the Foundry Skill, the VS Code toolkit and the plugin for the GitHub Copilot app, on Windows, macOS and Linux. Foundry Agent Service itself carries an explicit note on Learn: generally available, with some sub-features in public preview.
Those sub-features are where the planning happens. The Copilot harness is public preview with hard edges: bring your own model is unsupported, Foundry Memory is unavailable, and the service-managed sandbox exposes no CPU, memory or lifecycle settings. Voice agents, the A2A tool, computer use and browser automation are generally available. Insights and multi-turn evaluation are generally available, while annotations on traces, evaluation datasets, skill optimisation and cost optimisation are preview. Agent optimizer is one item where Microsoft's pages differ: the Book of News calls it generally available and the Foundry Agent Service post lists insights to agent optimizer as preview.
Regionally this layer suits Europe. Foundry projects can be created in Sweden Central, which the Agent Service region table also marks supported for agents, voice agents, computer use and private virtual networks. Where agent state lives depends on the setup. Basic setup keeps files, vector stores and conversations in Microsoft-managed storage and deletes conversations after 60 days. Standard setup puts them in your own Storage, Foundry IQ and Cosmos DB, which Microsoft describes as full control over data residency and access. That choice is the one we work through in Azure OpenAI integration engagements.
Running Foundry on your own Kubernetes is the one place where two Microsoft posts from the same day say different things. The AKS engineering post lists deploying open-weight models and running Foundry hosted agents on customer AKS clusters as limited preview, each behind a signup form. The Foundry Agent Service post lists deployment to customer-owned AKS as coming soon, with public preview targeted for early December.
Isolation: Agent Sandbox, Agent Grid, agent controls
Three preview services sit between an agent and what it can damage, and they do not overlap.
The Agent Sandbox cluster extension for AKS gives virtual-machine-level isolation through Kata VM isolation v2. Its Learn page is titled as preview and repeats the standard AKS warning that previews sit outside service-level agreements and are not meant for production. Requirements are narrow: v6 VM SKUs from nine named families, a region carrying AKS release 1025, and images only from Microsoft Container Registry or Azure Container Registry. The serverless runtime covers seven regions, Sweden Central among them. One detail in the walkthrough matters for European deployments: the workload space resource group must be created in West Central US, and the regional ARM profile routes requests to the West Central US management endpoint.
Agent Grid is the network layer, announced in preview, with Layer 3 to 7 policy and inspection for traffic among agents, MCP servers, models and external destinations. Its preview supports three execution options: ACI agent groups, AKS Agent Sandbox, and native Kubernetes workloads on AKS. The blog publishes no region list.
Agent controls in Foundry is a managed policy service in preview, built on the open-source Agent Control Specification and Agent Hooks. It evaluates a proposed action at one of eight intervention points and returns allow, deny or transform. Read its enforcement boundary section before drawing it as a wall on a diagram. Microsoft's words: agent controls "doesn't run your agent, execute its tools, sandbox its process, or control a path that bypasses your interception code". The host decides whether to obey.
Security and detection: Perception and Defender
Microsoft Perception is generally available, and the per-agent breakdown matters more than the headline. Alert Triage, Attack Investigation, Detection Authoring, Threat Intelligence, Recon and Posture Prioritization are generally available. Risk Reduction, Web App Pen Test and vulnerability shielding are preview, and AI readiness initiatives in Perception are stated for January 2027. The Learn application card adds a standing hedge that some capabilities may be in preview.
Perception is also the clearest case of two layers genuinely joining. The application card states that during setup Microsoft Entra creates a dedicated Agent ID for the agent, and that the permissions assigned to that Agent ID govern what data the agent can access. The identity layer is load-bearing here, not decorative.
Local agent discovery in Defender for Endpoint is generally available on supported Windows and macOS devices, except for scenarios the support matrix marks preview: the macOS VS Code extension entries, and agents under Windows Subsystem for Linux, which explicitly does not imply native Linux support. The matrix lists roughly 70 agents per platform plus their MCP servers. One prerequisite line matters in regulated Europe: the environment must be in the commercial cloud, and sovereign and national clouds are not supported. Plan 2 is the minimum licence for inventory and hunting, while risk levels and recommendations need Microsoft 365 E7 or Agent 365.
The inventory is queryable, which makes it the cheapest useful thing in this layer:
AgentsInfo
| where Platform == "LocalAgents"
| summarize arg_max(Timestamp, Name, Version, LifecycleStatus, RawAgentInfo) by AgentId
| where LifecycleStatus !in~ ("Deleted", "Uninstalled")
Runtime protection for local agents is the sharpest contradiction of the week. The Agent 365 Ignite blog says it is "Now generally available on Windows". The Defender for Endpoint support matrix on Learn, updated 13 November, heads its runtime protection section "Windows (preview)". The methods are documented either way. Agent-native event inspection uses vendor hook interfaces and covers Claude Code, GitHub Copilot CLI, the Copilot app, Claude Desktop and the VS Code extensions. Network inspection covers agents talking to well-known inference services, and does not cover certificate pinning or HTTP/3.
Operations: Azure SRE Agent
Azure SRE Agent reached general availability on 10 March 2026. Learn lists 22 deployment regions including Sweden Central, France Central, Italy North, Spain Central and UK South. Each agent deploys to one region and cannot be moved, but the region only determines where agent compute runs: with the right managed identity permissions it manages resources in any region or subscription. For anyone drafting a residency statement, that distinction is the one to get right.
Everything announced at Ignite sits on top in preview. Teams integration, Entra Agent ID, Agent 365 integration covering registry, Defender and Purview, agent collaboration over A2A, the AWS and GCP connectors, customer-managed key encryption and the Azure CLI are public preview. Slack integration is private preview. The one generally available operations agent in the stack is wired to the generally available identity and management planes by public preview integrations only.
Data governance: Purview and your own keys
Purview treats an Agent 365 agent instance largely as it treats a user, marking ten capability areas as supported for AI interactions, from posture management and auditing through labels, data loss prevention, insider risk and eDiscovery. One row is unsupported: encryption without sensitivity labels. Two footnotes qualify the rest. Content newly created by Agent 365 inherits no sensitivity label from its source items, so it is neither labelled nor encrypted automatically. And because an agent instance is unaware of a DLP block, Microsoft tells the agent owner to monitor the policy and the effect on downstream workflows. A control that fires silently against a non-human actor needs a human watching it.
Timing on the new Purview work is contradictory across two blogs from the same week. The Agent 365 post says Purview data loss prevention "can now detect and block sensitive information before Microsoft agents and key local agents process, expose, or transmit it", labelled now in preview. The Purview post labels "Protect the sensitive data local agents use" as public preview January 2027, alongside three other January-dated items.
Customer-managed keys are not a stack-wide feature. They are public preview in Azure SRE Agent. In Foundry Agent Service the nearest equivalent is standard setup, which moves agent state into Azure resources you own and key yourself. Agent 365 encrypts connected-platform credentials with Microsoft standard encryption and states that it does not currently support Advanced Data Residency.
The framework: Microsoft Agent Framework
Agent Framework reached version 1.0 for .NET and Python, described as production-ready with stable APIs and long-term support. It is the declared successor to Semantic Kernel and AutoGen and supports Foundry, Anthropic, Azure OpenAI, OpenAI and Ollama as providers. Learn marks Agent Framework for Go as public preview, requiring Go 1.26 or later, with declarative agents, CodeAct and functional workflows not implemented in Go at all.
The 1.0 announcement lists what the stability commitment excludes, and that list is exactly what connects the framework to the rest of the stack: Foundry hosted agent integration, Foundry tools, memory, observability and evaluations, Skills, the agent harness, DevUI, and the GitHub Copilot and Claude Code SDK integrations. Learn adds that managing whether data flows outside your organisation's Azure compliance and geographic boundaries is your responsibility when third-party systems are involved. That is the right place for the decision and a surprising place to find it.
One supply-chain note belongs here: the agent-framework PyPI package was briefly published in a malicious version carrying a credential stealer, with audit steps in our Azure audit for that compromise.
Status at a glance
Statuses are as stated on the pages cited at the end.
| Layer | Component | Stated status | What it does not reach |
|---|---|---|---|
| Identity | Entra Agent ID platform | Available to all Entra customers | Entra security features for agents need Agent 365 |
| Identity | Agent identity lifecycle automation | Generally available | Permission remediation and access packages are preview |
| Identity | Agent ID for discovered local agents | Preview | Depends on Defender discovery, commercial cloud only |
| Management | Agent 365 and agent registry | Generally available, Commercial, per user | No Advanced Data Residency, 30-day observability retention |
| Management | Connected platforms (7 vendors) | Read-only sync; 2 vendors observability "coming soon" | Manual sync only, deletions not reconciled, keys required |
| Management | Action Center, external tool catalogs | Preview in December, private preview in December | Not available for a Q4 decision |
| Build | Foundry Dev Pack, Foundry CLI | GA 18 November, GA 5 November | Foundry Canvas still a preview component |
| Build | Foundry Agent Service | Generally available, sub-features preview | Azure AI Agents unsupported in Azure Government |
| Build | Evaluation loop | Insights and multi-turn eval GA; optimisers preview | Agent optimizer status differs between two MS pages |
| Build | Foundry agents on customer AKS | Limited preview, or public preview early December | Allowlist signup, no region list |
| Isolation | Agent Sandbox extension for AKS | Preview, excluded from SLAs | Workload space resource group must be West Central US |
| Isolation | Agent Grid | Preview | ACI, AKS Sandbox and native AKS only; no managed Foundry |
| Isolation | Foundry agent controls | Preview | Returns a verdict; your host must enforce it |
| Detection | Microsoft Perception | GA, with 2 agents preview and a Learn hedge | AI readiness initiatives stated for January 2027 |
| Detection | Defender local agent discovery | GA on Windows and macOS, WSL preview | Commercial cloud only; no sovereign or national clouds |
| Detection | Runtime protection for local agents | GA on Windows (Agent 365 blog) or preview (Learn) | No cert-pinned or HTTP/3 traffic; Windows only |
| Operations | Azure SRE Agent | GA since 10 March 2026, 22 regions | Identity, Agent 365, CMK and cloud connectors all preview |
| Data | Purview for Agent 365 | Supported across 10 capability areas | No encryption without labels; new content inherits no label |
| Data | Purview DLP for local agents | Preview now (Agent 365 blog) or January 2027 (Purview blog) | The endpoint path is the January-dated one |
| Framework | Microsoft Agent Framework 1.0 | 1.0 for .NET and Python; Go public preview | Foundry hosting, observability and Skills are preview |
Where the stack does not join up
Six gaps are structural rather than temporary, in the sense that closing them needs a product decision and not just a GA date.
Identity is free, acting on identity is not. Every Entra customer can mint agent identities. Applying Conditional Access, Identity Protection, governance and network controls to them requires Agent 365, which means E7 or an add-on on top of E5. The architecture diagram shows one identity layer. The purchase order shows two.
The build layer does not hand identities to the run layer. Copilot Studio, Security Copilot and Foundry create Agent IDs through documented product integrations. An agent you write with Agent Framework and host on AKS or App Service gets an identity because you called Graph, used the Entra Auth SDK sidecar, or configured workload identity federation. Agent Framework 1.0's stability commitment does not mention Entra Agent ID at all.
Policy evaluation and policy enforcement live in different layers. Foundry agent controls returns a verdict and says it does not run your agent or control a path that bypasses your interception code. Agent Grid enforces in Azure's network infrastructure, outside the agent's control, but only for ACI agent groups, AKS Agent Sandbox and native AKS workloads. Agents in the managed Foundry Agent Service and agents on developer laptops appear in neither list. The enforcement guarantee depends on where the agent runs, and the first place most organisations will run is the place Agent Grid does not cover.
The third-party registry is a snapshot, not a feed. Connected platforms is read-only, synchronises only when an administrator clicks, and does not guarantee that an agent deleted in Bedrock or Vertex disappears from the registry. Every supported platform needs a key-based credential. Microsoft's pages conflict here: the Agent 365 Ignite blog describes secretless authentication using Entra workload identity federation as a preview capability, while Learn says keyless options such as workload identity federation or role assumption are not currently supported. Either way, the governance plane for non-Microsoft agents runs on stored secrets, one layer above an identity plane whose selling point is removing them.
Detection stops at the cloud boundary, and so does the sovereign story. Defender local agent discovery requires the commercial cloud and does not support sovereign or national clouds. Foundry's region page lists Azure AI Agents, the agents playground and fine-tuning among features unsupported in Azure Government. An organisation that chose a sovereign cloud for regulatory reasons keeps the identity layer, loses local-agent detection, and loses the managed agent runtime. Nothing at Ignite changed that.
There is no single residency assertion for an agent. Four layers answer differently. Agent 365 observability stores customer content in the tenant's default geography, replicates some service-generated data to the EU for EU tenants and to the US otherwise, honours EU Data Boundary commitments and does not support Advanced Data Residency. Foundry Agent Service endpoints are regional with data in the endpoint's region, unless you choose basic setup and get Microsoft-managed storage. Perception preselects a GPU evaluation location from the tenant's Customer Data storage location. Azure SRE Agent runs compute in one region while acting anywhere you grant permission. Covering all four in one record of processing activities means reading four Learn pages that do not reference each other.
Retention is misaligned too. Agent 365 keeps observability data 30 days. Foundry basic setup deletes conversations after 60 days and returns 404 for expired ones. Neither window matches a 90-day forensic question, so export before the clocks run out.
What this means for an adoption decision
Adopting Microsoft's agent platform today gets you a real inventory, an identity object per agent, a generally available developer toolchain with Sweden Central support, a generally available multi-agent SOC and a generally available operations agent. That is enough to run a governed pilot.
Four things are worth settling before a 2027 roadmap depends on it. Confirm whether your tenant qualifies for the Agent 365 or E7 licence that turns identity into enforcement, because the rest of the governance story assumes it. Decide where agents execute, since isolation and network policy exist only on the self-hosted path and the sandbox preview still anchors a resource group in West Central US. Pick standard setup in Foundry Agent Service if residency is contractual rather than a preference. And treat every cross-layer integration as preview until the specific Learn page says otherwise.
subscribe # AI-nyheterna som spelar roll, utan bruset
Sources
- Microsoft Ignite 2026 Book of News
- Microsoft Entra at Ignite 2026
- What is Microsoft Entra Agent ID?
- How are agent identities created?
- What's new in Microsoft Agent 365 at Ignite 2026
- Microsoft Agent 365 overview
- Connected platforms in Microsoft Agent 365
- Data handling, residency and compliance in Agent 365 observability
- The Foundry Dev Pack is here
- What's new in Foundry Agent Service at Ignite 2026
- Quotas, limits and regions for Foundry Agent Service
- Foundry feature availability across cloud regions
- What are agent controls? (preview)
- Run models and agents on your own AKS clusters
- Agent Sandbox cluster extension (preview)
- Introducing Agent Grid
- Security for the agentic era, powered by Microsoft Perception
- Application card for Perception
- Discover local AI agents with Defender for Endpoint
- Defender for Endpoint AI agent support matrix
- Announcing general availability for Azure SRE Agent
- What's new in Azure SRE Agent at Ignite 2026
- Supported regions for Azure SRE Agent
- Secure the data that fuels AI with Microsoft Purview
- Use Purview for data security and compliance for Agent 365
- Microsoft Agent Framework overview
- Microsoft Agent Framework version 1.0