Microsoft Agent Framework 1.0: AutoGen and Semantic Kernel Finally Unified
Microsoft Agent Framework has hit 1.0 General Availability — the production-ready convergence of AutoGen and Semantic Kernel into a single, coherent framework for building multi-agent systems on Azure. This is open-source, available in both .NET and Python, and ships with native support for both Agent-to-Agent (A2A) and Model Context Protocol (MCP) for cross-runtime interoperability.
For the enterprise teams I work with, the fragmentation between AutoGen and Semantic Kernel has been a persistent source of confusion and duplicated effort. 1.0 GA resolves that. Here is a technical breakdown of what ships, what it replaces, and the honest architectural trade-offs.
The 3-Layer Architecture
Agent Framework 1.0 is structured across three layers that separate concerns cleanly:
- Framework layer: The programming model — agent definitions, orchestration patterns, tool registration, memory interfaces. This is what developers interact with in Python or .NET code.
- Runtime layer: Foundry Agent Service on Azure — managed execution, sandboxed sessions, persistent state, scaling, and observability. You can also self-host the runtime on AKS or other infrastructure.
- Connectivity layer: A2A for agent-to-agent communication across runtimes and organisations. MCP for tool and data source integration. These protocols make Agent Framework interoperable with the broader ecosystem.
The Four Orchestration Patterns
Agent Framework ships with four built-in orchestration patterns that cover the majority of multi-agent use cases:
- Sequential: Agents execute in a defined order, each consuming the output of the previous. Suitable for pipeline workflows with clear stage dependencies.
- Concurrent: Multiple agents execute in parallel on independent subtasks. Suitable for fan-out workloads where subtasks do not depend on each other.
- Group chat: Multiple agents participate in a shared conversation, with a moderator agent managing turn-taking and synthesis. Suitable for deliberation and review workflows.
- Handoff: An orchestrator agent delegates to specialist agents based on task type, with explicit handoff protocols. Suitable for routing workflows where different task types require different expertise.
What Foundry Agent Service Handles For You
When paired with Azure Foundry Agent Service, five operational concerns that you would otherwise build yourself are managed by the platform:
- Sandboxed session execution with automatic isolation between agent runs
- Persistent state management across sessions using the 3-tier memory model
- Tool execution with permission enforcement and audit logging
- Scaling and capacity management across concurrent agent sessions
- Integrated observability through Azure Monitor with agent-specific metrics
A2A and MCP Cross-Runtime Interoperability
A2A (Agent-to-Agent protocol) enables agents built with Agent Framework to communicate with agents built on other runtimes — including AWS Bedrock Agents, Google ADK agents, and agents built with LangGraph. This is important for enterprise deployments where different teams or organisations may be running different agent frameworks and need to compose capabilities across them.
MCP integration allows Agent Framework agents to consume tools and data sources exposed by any MCP server, without custom integration code. The combination of A2A and MCP makes Agent Framework a viable orchestration layer for heterogeneous multi-agent architectures.
Honest Trade-offs
- Azure dependency: The full value of Agent Framework requires Foundry Agent Service — self-hosting the runtime is possible but loses managed capabilities
- Migration cost: Teams with existing AutoGen or Semantic Kernel code face a migration effort to the unified 1.0 API surface
- LangGraph competition: LangGraph's graph-based orchestration model offers more fine-grained control for complex agentic workflows and has a significant community
- A2A ecosystem maturity: Cross-runtime A2A interoperability is early-stage — the protocol is sound but real-world multi-vendor deployments will surface edge cases
Key Takeaways
- Microsoft Agent Framework 1.0 unifies AutoGen and Semantic Kernel into a single production-ready framework — the fragmentation era is over
- Four built-in orchestration patterns cover sequential, concurrent, group chat, and handoff workflows out of the box
- A2A and MCP enable cross-runtime interoperability with other agent frameworks and tool ecosystems
- When paired with Foundry Agent Service, operational concerns including state, scaling, and observability are managed by the platform
- Teams already invested in LangGraph or AWS Bedrock Agents should evaluate migration cost carefully — the competitive landscape is stronger than it was 12 months ago
