What is Microsoft Agent Framework?
Microsoft Agent Framework is Microsoft's open-source Python and .NET development kit for building AI agents and typed, graph-based workflows. Microsoft identifies it as the direct successor to AutoGen and Semantic Kernel, with separate abstractions for agents, sessions, context, middleware, tools, and workflow execution.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Microsoft Agent Framework separates open-ended agent behavior from explicit orchestration. An Agent combines a model client, instructions, tools, and optional context providers; an AgentSession carries conversation state; middleware can inspect or constrain calls; MCP clients expose external capabilities. For multi-step systems, a Workflow connects typed executors with edges, routing, nesting, checkpointing, and request-response pauses for human input. That design is a migration, not a namespace change from AutoGen: teams must remap model clients, message types, tools, state, and team patterns. Checkpoints preserve workflow state, but they do not make database writes, payments, or messages exactly-once; production systems still need idempotency keys, effect records, reconciliation, authorization, and tenant isolation.
Key Characteristics
- Dual SDK scope: provides related agent and workflow abstractions for Python and .NET
- Agent composition: combines model clients, instructions, tools, sessions, context providers, and middleware
- Typed workflows: routes messages through graph edges and executors instead of relying only on free-form agent conversation
- Operational controls: supports checkpointing, request-response pauses, human interaction, and telemetry integration
- Protocol integration: can connect agents to MCP servers while leaving trust, permission, and data-boundary decisions to the application
Common Use Cases
- Building a Python or .NET assistant with session state and governed tools
- Composing deterministic functions and model-driven agents in one typed workflow
- Pausing a long-running process for human approval and resuming from stored state
- Migrating an AutoGen or Semantic Kernel application with explicit API mapping
- Adding middleware, telemetry, and MCP integrations around enterprise agent execution
Example
Loading code...Frequently Asked Questions
Is Microsoft Agent Framework the replacement for AutoGen?
Microsoft describes Agent Framework as the direct successor to both AutoGen and Semantic Kernel for new development. AutoGen remains a separate project and API. Existing applications do not migrate by renaming imports; clients, messages, tools, sessions, state, team patterns, and observability hooks need explicit mapping and regression tests.
How is Microsoft Agent Framework different from AG2?
Microsoft Agent Framework is developed by Microsoft as the successor path for AutoGen and Semantic Kernel and provides Python and .NET SDKs. AG2 is a separate project that continues the classic `autogen` and ConversableAgent lineage and has its own roadmap. Their packages and APIs are not interchangeable.
When should I use an Agent instead of a Workflow?
Use an Agent when the task is underspecified and benefits from model-directed tool selection or conversation. Use a Workflow when steps, routing, parallelism, approvals, or recovery boundaries must be explicit. A Workflow can include agents alongside deterministic executors, so the two abstractions are complementary.
Does workflow checkpointing guarantee exactly-once execution?
No. A checkpoint records framework state from which execution can resume; it cannot prove whether an external payment, email, deployment, or database write committed before a crash. Consequential tools need idempotency keys, durable effect records, reconciliation, and compensation where reversal is possible.
What should teams verify before adopting Microsoft Agent Framework?
Check the current official release status, exact Python or .NET package versions, provider support, migration coverage, storage adapters, telemetry, and deployment constraints. Then test task success, unsafe tool attempts, interruption and resume behavior, latency, cost per successful run, and rollback against a pinned evaluation set.