Microsoft Agent Framework 1.0: SDK Strong, Context Gap Remains

What happened

Microsoft has shipped version 1.0 of Agent Framework, which it describes as an open-source SDK and runtime for building AI agents and multi-agent workflows on enterprise-grade foundations 1. General availability arrived in April 2026, and the release consolidates two of Microsoft's existing agent efforts into one product 2. Microsoft's own write-up organizes the framework around two primary categories of capability, positioning it as both a developer toolkit and an execution environment rather than a single library 1.

The consolidation is the most important part of the launch. Agent Framework combines AutoGen's simple abstractions for defining agents with the enterprise-oriented features that Semantic Kernel had accumulated 2. For developers who had to choose between AutoGen's ease of experimentation and Semantic Kernel's production focus, Microsoft is now offering one answer.

Why the merger matters

Microsoft's agent tooling has had a fragmentation problem. AutoGen grew out of research and appealed to people prototyping multi-agent conversations. Semantic Kernel served teams embedding AI into business applications. Running two overlapping frameworks split documentation, community attention, and roadmap credibility. Folding them into a single SDK signals that Microsoft wants one supported path for agent development, which matters to enterprises that weigh long-term vendor commitment as heavily as features.

Microsoft also handled the transition with some care. Semantic Kernel v1.x will continue to receive critical fixes for at least a year after Agent Framework's April 2026 GA 2. That window gives teams with Semantic Kernel code in production time to plan a migration instead of being forced into one. The practical takeaway is that Agent Framework is where new investment should go, while existing Semantic Kernel deployments get a defined, though limited, runway.

Interoperability as a selling point

Agent Framework's protocol support is broad. It works with the Model Context Protocol (MCP), Agent-to-Agent (A2A), AG-UI, and OpenAPI, so agents can call external tools and coordinate across runtimes and vendors 2. This reflects where the industry has gone. Agent systems are rarely built with a single vendor's components, and frameworks that lock developers into a closed tool ecosystem lose ground to ones that adopt shared standards.

MCP covers tool and data access. A2A addresses communication between agents, including agents built on other stacks. AG-UI targets the connection between agents and user-facing interfaces. OpenAPI support makes the large existing body of REST services available to agents. Taken together, these choices suggest Microsoft wants Agent Framework to act as an orchestration hub in heterogeneous environments, not a walled garden.

Where the sources diverge

The two accounts agree on what the framework is but differ in emphasis. Microsoft's announcement presents it as an enterprise-ready foundation for agentic AI 1. Atlan's analysis grants the engineering strength but questions the enterprise claim, framing the release as a strong SDK with a shaky enterprise fit 2.

Atlan's argument is that orchestration, meaning how agents are sequenced, coordinated, and connected to tools, is now largely solved inside the SDK. Governed business context is not 2. By this reading, the deciding factor for production accuracy is not how well agents pass messages to each other. It is whether they work from trustworthy, well-defined, access-controlled organizational knowledge 2.

Atlan sells data governance and context tooling, so it has a commercial reason to stress this gap. That doesn't make the point wrong. It does mean readers should treat it as a perspective from one side of the market and not as a neutral benchmark.

Our reading

Agent Framework 1.0 looks like a meaningful cleanup of Microsoft's agent strategy. Merging AutoGen and Semantic Kernel removes a long-standing source of confusion. The defined support window for Semantic Kernel respects existing users. The commitment to MCP, A2A, AG-UI, and OpenAPI positions the framework well in a multi-vendor world.

The critique of enterprise fit still points at something real. A mature SDK makes agents easier to build and wire together, but it does not by itself tell an agent what a "customer," "active contract," or "approved discount" means inside a particular company. Many enterprise agent failures are likely to come from ambiguous or ungoverned data, not from orchestration bugs. Organizations evaluating Agent Framework should treat it as solid plumbing and budget separately for the semantic and governance layer that sits underneath their agents.

In short, Microsoft has largely solved the problem of building agents, while the harder problem of making agents reliably correct about the business still belongs to each organization.