AzureFeeds

May 15, 2026  |  8 min read 

If you’ve been paying attention to Microsoft over the past six months, you’ve noticed a recurring theme: agents are everywhere. They’re reasoning over your SharePoint folders, reconciling your accounts, and if you squint at the Ignite 2025 keynote, they’re about three sprints away from filing your taxes. 

Developers need to debug, evaluate, monitor, and improve agents before and after deployment. IT and security teams need to know which agents exist, who owns them, what they can access, and whether they comply with organizational policy. 

That’s why Microsoft has 2 complementary governance surfaces: Microsoft Foundry Control Plane and Microsoft Agent 365. Foundry Control Plane is primarily for the teams that build and operate agents. Agent 365 is primarily for the teams that govern agents. This post breaks down what each one does, why most enterprises will end up using both. 

TL;DR: Use Microsoft Foundry to build, evaluate, trace, deploy, and operate agents. Use Microsoft Agent 365 to discover, govern, secure, and manage agents across the organization. Most enterprises will need both.

Microsoft Foundry: The Platform Where Agents Are Born 

Microsoft Foundry (formerly known as Azure AI Foundry and Azure AI Studio) is the end-to-end platform for building and operating AI agents at enterprise scale It hosts over 11,000+ models including GPT-5.5, Claude Opus 4.8, Grok 4.3, and so many open-source models. It has a managed agent runtime called Foundry Agent Service. It bundles agentic RAG through Foundry IQ, secure tool connectors, edge deployment through Foundry Local, and the star of today’s post: the Foundry Control Plane. 

 
Foundry Control Plane: The Developer’s Governance Layer 

If Foundry is the factory, the Control Plane is the quality control department, the security team, and the operations dashboard rolled into one. It was announced at Ignite 2025 and reached major GA milestones in March 2026, when evaluations, monitoring, and tracing all became generally available. It’s built specifically for developers and platform engineers who need to understand what’s happening inside every agent from code to production. 

Fleet Management. You get a single view of every agent across every project in your subscription. Agents built natively in Foundry appear automatically with an Entra Agent ID. External agents built with LangChain, LangGraph, or the OpenAI SDK etc. connect through the AI Gateway and show up in the same fleet view. 

Guardrails and Controls. These operate at the agent level, as well as at model deployments level. They cover prompt injection mitigation, groundedness verification, sensitive data detection, protected materials, and a newer addition called task adherence, which catches agents drifting off course. Think of task adherence as the detector that flags when your AI intern decides to reorganize the supply closet instead of finishing the quarterly report. Controls now also extend to tool calls and tool responses, which means you can govern what agents do when they interact with external systems. 

Observability (GA). Evaluations, monitoring, and tracing are deeply integrated with Azure Monitor. You get out-of-the-box evaluators for coherence, relevance, groundedness, and safety, along with custom evaluators in preview for business-specific assessments. Auto-instrumentation works across Semantic Kernel, LangChain, LangGraph, OpenAI Agents SDK and more. Your AI incidents route through the same on-call rotations and alerting systems you already use for everything else. 

Continuous Evaluation. This is the real differentiator. Most evaluation tools only work as a pre-deployment gate. Foundry’s evaluators run identically at dev time, in CI/CD, and against live production traffic. You get one consistent measurement system across the entire lifecycle instead of two disconnected ones that can’t be compared. 

Security. Native integration with Defender for Cloud, Purview, and Entra, plus expanded partner coverage from Palo Alto’s Prisma AIRS and Zenity. 
 
Developers access Microsoft Foundry through the Foundry portal at https://ai.azure.com or through Azure-native tools such as SDKs, REST APIs, Azure CLI, Bicep, and CI/CD pipelines. After signing in with an Entra ID account that has access to the Azure subscription and Foundry project, developers can create agents, choose models, connect tools, add knowledge sources, configure evaluations, trace runs, and monitor production behavior. This is why Foundry is best positioned as the builder platform: it gives engineering teams the workspace to design, test, deploy, and operate agents 

Agent 365: The IT Admin’s Answer to Agent Sprawl 

While Foundry Control Plane is focused on the developer experience, Microsoft Agent 365 serves a different audience. It’s for IT administrators, security teams, and business leaders who need to manage agents at the organizational level. Agent 365 is not an agent-building platform. IT admins can use Agent 365 as the control plane for governance, security, identity and lifecycle management. 

Agent 365 became generally available on May 1, 2026 at $15 per user per month as a standalone license. IT admin needs to log into M365 Admin Center, and there, they will see a new tab on the left pane, called ‘Agents’. Same portal https://admin.microsoft.com/ 

The Core Idea: Treat Agents Like Users 

Microsoft’s philosophy here is refreshingly simple. Agents are digital teammates. They need identity management, security controls, and compliance governance that you already apply to your human employees. Every agent gets its own Microsoft Entra Agent ID, an immutable identity record that works just like a user or app registration. No new governance model to learn. You just extend what you already have. 

Agent Registry. This is a unified inventory of every agent in your tenant. It covers agents built with Microsoft AI platforms, agents from ecosystem partners, agents you register manually, and shadow agents (the unauthorized ones that your employees spun up without IT knowing about it). Think of it as Active Directory for AI agents. 

Access Control and Identity Governance. Conditional Access evaluates agent context and risk in real time before granting access. Entitlement management ensures agents only have the permissions they actually need, for only as long as they need them. If an agent’s owner leaves the company, governance workflows make sure the agent doesn’t become an orphaned, over-privileged ghost haunting your directory. 

Threat Protection. Defender detects anomalous agent behavior in real time, and Identity Protection flags agents that suddenly start accessing unfamiliar resources or making an unusual number of sign-in attempts. It’s the AI equivalent of someone badge-swiping into the server room at 3 AM. 

Data Compliance. Purview applies DLP, Insider Risk, Communication Compliance, eDiscovery, and Data Lifecycle Management to agent interactions. The same compliance toolchain you use for people now covers your digital workforce. 

Network Controls and Shadow Agent Detection. Full visibility into all agent network activity, web categorization for APIs and MCP servers, prompt injection detection at the network level, and perhaps most importantly, the ability to surface unauthorized agents that employees have created without IT approval. Better to find them through governance than through a breach. 
 
With that being said, Agent 365 is not only an admin portal. Microsoft also provides an Agent 365 SDK and Agent 365 CLI for developers who need to bring custom or third-party agents into the Agent 365 governance model. The Agent 365 SDK layers enterprise capabilities on top of agents built with any frameworks: it gives agents Entra-backed agent identity, governed MCP tool access, OpenTelemetry-based observability, notifications, and lifecycle governance through Agent 365. The Agent 365 CLI is the command-line tool for building, configuring, deploying, and publishing Agent 365-ready applications. It can help create agent blueprints, configure MCP servers and permissions, deploy agent code to Azure App Service, package the app manifest, and publish the package for Microsoft 365 admin center approval. 

The Comparison Table 

Who Needs What? 

Foundry Control Plane is for you if you’re building custom agents, need framework-agnostic observability, care about tracing and evaluations, and want token-level cost analytics. 

Agent 365 is for you if you’re governing agents across an organization, need a centralized registry that includes shadow agents, want conditional access and compliance policies applied to agents, and care about threat detection. 

Both make sense if you’re a regulated enterprise deploying agents at scale. Developers build in Foundry with Control Plane quality gates, then publish to Agent 365 where IT handles the organizational governance. This is the pattern Microsoft is designing for. 

Best Practices 

Begin with identity. Every agent should get an Entra Agent ID from day one. No generic app identities, no shared credentials, assign a clear owner, and make sure agent access can be reviewed, disabled, and audited like any other enterprise identity. 

Evaluate continuously. Run Foundry evaluators at dev time, in CI/CD, and against production traffic. Model updates and prompt drift will catch you off guard if you only test before deployment. 

Apply least-privilege. Use Agent 365’s entitlement management for time-bound, scoped permissions. Agents should not accumulate access over time like a digital hoarder. 

Hunt shadow agents early. Turn on detection before your employees’ creativity outpaces your governance policies. 

Layer your guardrails. Foundry guardrails handle behavior quality. Agent 365 controls handle security posture. Both layers are necessary. Neither is sufficient on its own. 

Conclusion 

Microsoft’s agent governance architecture is best understood as a separation of responsibilities. 

Foundry Control Plane is the builder control plane. It helps teams understand how agents behave, how they perform, what they cost, and whether they meet quality and safety expectations. 

Agent 365 is the enterprise admin control plane. It helps IT and security teams discover, govern, secure, and manage agents across the organization. 

The two platforms solve different problems for different audiences. Together, they give Microsoft a credible end-to-end governance model: engineering control in Azure, organizational control in Microsoft 365 

Now if you’ll excuse me, I need to go register all the shadow agents my team spun up while I was writing this. 

Resources 

Foundry Control Plane: learn.microsoft.com/en-us/azure/foundry/control-plane/overview 

Agent 365: learn.microsoft.com/en-us/microsoft-agent-365/overview 

Foundry Portal: https://ai.azure.com/  

Get Started

For developers:

• Quickstart: Build your first agent in Foundry (15 min)


• Configure Control Plane guardrails


• Set up continuous evaluation

For IT admins:

• Quickstart: Enable Agent 365 in your tenant (10 min)


• Discover shadow agents in your organization


• Configure Conditional Access for agents

Community:

• Join the Microsoft 365 Copilot Tech Community for agent governance best practices and Q&A.