Microsoft's enterprise-ready multi-agent framework: stable APIs, long-term support, multi-provider models, and cross-runtime orchestration via A2A and MCP.
Before you start
Python 3.10+ • pip install microsoft-agent-framework • API keys for at least one model provider (Anthropic, OpenAI, or Google). If you have existing AutoGen code, the framework provides migration guides.
Agent Framework separates agents (stateless reasoning units) from runtimes (execution environments that manage message passing and lifecycle). This split lets you run the same agent logic locally during development and on Azure or another cloud runtime in production without changing agent code.
Agent Framework uses typed message schemas, agents declare what messages they send and receive. This makes agent boundaries explicit and prevents the loose string-passing that makes AutoGen workflows hard to debug at scale.
A2A (Agent-to-Agent) lets agents running in different runtimes, or built with different frameworks, communicate over a standard protocol. Wire your agents together by declaring their A2A endpoints, then test communication with a simple end-to-end task before adding complexity.
Agent Framework natively exposes agents as MCP servers and can consume external MCP tools. This means your agents can be called from Claude Code, Cursor, or any MCP client without additional wiring.
Core Agent Framework concepts and API patterns for building production multi-agent systems.
| Command / Skill | What it does | When to use it |
|---|---|---|
Agent class | Stateless reasoning unit, declares input/output schemas and uses a configured LLM | Base building block for every workflow, one agent per distinct role or capability |
Runtime | Execution environment managing agent lifecycle and message passing, local for dev, Azure/cloud for production | Choose local runtime during development; swap to cloud runtime without changing agent code |
A2A protocoladvanced | Agent-to-Agent communication standard enabling cross-runtime and cross-framework agent coordination | Connecting agents in different runtimes, or integrating with agents built on other frameworks |
MCP integrationadvanced | Expose any agent as an MCP server, or consume external MCP tools from within agents | Making agents callable from Claude Code, Cursor, or any MCP-compatible client |
Multi-provider model config | Configure Claude, GPT, Gemini, or any provider per agent, swap via config, not code | Routing different agents to different models by cost, capability, or latency requirements |
Human-in-the-loop checkpoints | Pause execution and request human approval before proceeding to next agent or irreversible action | Any workflow step that deploys, sends, or permanently modifies data |