Microsoft Agent Framework 1.0: SDK Unified, Stack Still Split

What Microsoft shipped

Microsoft has released version 1.0 of its Agent Framework, dated April 3. The release brings together two open-source SDKs that previously could not work with each other, giving developers one toolkit for building AI agents 2. In practice, the framework combines the lightweight agent abstractions of AutoGen with the enterprise-oriented features of Semantic Kernel 3. Teams that have already built on Semantic Kernel are not being cut off. Its v1.x line will continue to get critical fixes for at least a year after the framework's general availability in April 2026 3.

The framework also supports several interoperability standards: MCP, A2A, AG-UI and OpenAPI. That lets agents call external tools and coordinate with other runtimes and vendors 3. For developers who have been switching between Microsoft's two agent SDKs, this consolidation fixes a real problem.

On the governance side, Microsoft is offering Agent 365. It gives each agent an Entra Agent ID, an entry in a registry, and tenant-wide observability through a single console 1. It costs $15 per user per month on its own, or comes included in the Microsoft 365 E7 bundle at $99 1. Defender and Purview enforcement also extend to these non-human identities. That targets the "shadow agent" visibility gap many enterprises say they have 1.

A unified SDK is not a unified stack

The three sources mostly agree on what Microsoft has delivered. They differ on how much it matters. Forbes argues that the 1.0 release covers only the build layer of a stack that enterprises have to assess in three separate tiers [2]:

• the SDK their engineers write against
• the runtime that hosts agents in production
• the governance layer that handles identity and compliance at scale

In Forbes' view, Microsoft's agent offering still spreads across too many surfaces for building, deploying, governing and distributing agents. Google Cloud and AWS, by contrast, each offer a clearer default path from framework to managed runtime 2. The advice to executives evaluating Azure is to decide which surface their teams will standardize on at each layer. They should also budget for the organizational cost of making those decisions 2.

Agent 365 fills in the governance tier, but it does not join that tier to the others. The license does not cover agent runtime costs 1. External agents that are never registered stay invisible unless someone writes policy to bring them in 1. This supports the fragmentation argument: governance is a separate product with separate billing, and its coverage depends on what organizations choose to register.

The context problem underneath

Atlan makes a narrower point. The framework handles orchestration inside the SDK, but it does not provide governed business context, and Atlan says that gap determines how accurate agents are in production 3. An agent can be well coordinated and still give wrong answers if the data it uses is poorly defined, outdated or ungoverned.

The Agent 365 material describes a related tension. Purview label policies need human judgment to balance safety against usefulness. Labels that are too strict can deprive useful agents of the context they need. Controls that are too loose bring the shadow-agent problem back inside an officially approved platform 1. The recommended approach is to pilot label policies on low-risk agents before applying them more widely 1.

Taken together, the two sources describe the same bottleneck from different directions. Atlan sees missing context as a data-layer gap. The Agent 365 material sees it as a policy-tuning problem. Either way, the hard part of enterprise agents now sits outside the SDK.

Reading the situation

The fair reading is that Microsoft has made real progress at the layer developers deal with first. Unifying AutoGen and Semantic Kernel, supporting open protocols like MCP and A2A, and committing to a long support window for existing code are all reasonable engineering decisions 3. Agent 365 also responds to a widely reported governance problem by giving agents identities and visibility instead of leaving them as anonymous scripts 1.

The build-run-govern split is still there, though. An enterprise adopting Microsoft's agent stack has to choose a framework, a hosting surface and a governance license separately, and then add its own work on data context and policy design. Forbes' comparison with Google Cloud and AWS points to the competitive risk: rivals that offer one obvious path can win projects even if their components are no better individually 2.

For buyers, this argues for treating Microsoft's agent offering as a set of parts to assemble, not a finished platform. The SDK question is largely settled. The costs of runtime, governance licensing and context engineering still need to be worked out before anyone commits.