Anyone who has worked on AI agent development has likely been frustrated by Microsoft's framework choices: Semantic Kernel and AutoGen each have their own APIs, different documentation styles, and non-overlapping community directions. To build a production-ready agent application, you either had to piece together frameworks or wait for Microsoft to unify its approach.
On April 3, the wait ended. Microsoft released Agent Framework 1.0, formally merging the two into a single SDK that supports both .NET and Python, and explicitly promises stable APIs and long-term support.
Semantic Kernel + AutoGen = Agent Framework
This merger is more than a simple rebranding. Microsoft explains that Semantic Kernel provides enterprise-grade infrastructure and service connectivity, while AutoGen offers flexible multi-agent orchestration concepts. Combining the two approaches covers the full spectrum from simple assistants to complex workflows.
Microsoft has stated that Semantic Kernel and AutoGen are now in maintenance mode, with all new features developed only in Agent Framework. Migration guides for developers already using these frameworks have been published simultaneously.
Compared to before, several areas have seen substantive changes:
- Multi-provider support: Built-in connectors for Azure OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, and Ollama, eliminating the need to write custom adapters
- Middleware pipeline: Insert content safety filters, logging, and compliance policies at various stages of agent execution without affecting the prompt itself
- Graph engine: Define workflows between agents and functions using graph structures, supporting deterministic processes and checkpoint recovery
- Pluggable memory: Built-in session history, key-value state, and vector retrieval, with backends supporting Redis, Neo4j, Mem0, or custom storage
On the protocol side, both MCP and A2A are supported. MCP is used to discover and invoke external tools, while A2A enables agents from different frameworks to collaborate with each other.
Getting Started with Five Lines of Code
Microsoft promotes a "5 lines of code, zero to agent" approach, with straightforward installation:
Python: pip install agent-framework
.NET: dotnet add package Microsoft.Agents.AI
Orchestration patterns support sequential, concurrent, handoff, group chat, and Magentic-One modes. YAML declarative configuration is also included, allowing agent definitions and workflow topologies to be managed under version control.
On the development tools side, DevUI is currently in preview. It is a browser-based debugging interface that visualizes agent execution flows, useful for troubleshooting deadlocks or logic issues in multi-agent scenarios.
What Problems Does Unification Solve?
Mitch Ashley, Vice President at Futurum Group, characterized the release as "a substantial advancement for enterprise AI development," highlighting the significance of multi-vendor support and compliance hooks for production deployment.
These two points are indeed critical. From an enterprise buyer's perspective, locking AI infrastructure to a single major vendor is both a compliance requirement and a risk management issue. From a developer's perspective, rewriting connection logic every time you switch model providers is a real friction cost.
Agent Framework 1.0 addresses this at the framework level: provider switching becomes a configuration item, requiring no changes to business logic.
Interestingly, Microsoft does not force you to use Azure. Anthropic Claude, Google Gemini, and Amazon Bedrock are all first-class citizens, and Ollama is also on the support list. This open attitude is somewhat unusual for Microsoft, likely reflecting an awareness that strong-arming would scare away customers.
How Does It Compete with LangChain and CrewAI?
The market is not short of agent frameworks: LangChain raised $125 million last year to become a unicorn, CrewAI added streaming tool call events, and the OpenAI Agents SDK is built directly into the API platform.
Microsoft has two differentiators: first, priority support for .NET developers—there is a substantial number of enterprise teams using C# and .NET that have not been prioritized by other frameworks; second, deep integration with Azure App Service and Microsoft Foundry, making migration costs low for workloads already running on Azure.
Additionally, support for the A2A (Agent-to-Agent) protocol is worth noting. A2A 1.0 is still in a "coming soon" state, but once stable, agents from different frameworks and vendors will be able to communicate directly—this is a foundational layer for agent interoperability.
The Significance of 1.0
In the AI toolchain, many things remain in preview for a long time, with production readiness in question. Microsoft's decision to apply the 1.0 label is a clear statement: this version is serious, stable APIs will not introduce breaking changes casually, and long-term support is promised.
For engineering teams evaluating agent development frameworks, this is at least an option worth serious consideration.
Sources: CocoLoop, Microsoft Agent Framework Version 1.0 (Microsoft Agent Framework Dev Blog, 2026-04-03); Microsoft Ships Agent Framework 1.0: A Production-Ready Foundation for Multi-Agent AI (Techstrong.ai)