What is Microsoft Agent Framework?
Microsoft Agent Framework is Microsoft's open-source agent and workflow platform; its Python and .NET core reached version 1.0, while the separately maintained Go SDK remains in public preview.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Separate stable core from preview surfaces
The April 2026 release made core Python and .NET agents, middleware, memory/context providers, workflows, and common orchestration patterns stable. DevUI, some hosted integrations, skills, and other extensions can retain preview status independently. The Go SDK is developed in a separate repository and remains public preview, with documented gaps such as declarative agents, some hosting features, and handoff orchestration.
Map AutoGen behavior before migrating
Microsoft AutoGen uses an event-driven Core and higher-level AgentChat teams; Agent Framework centers agents plus typed data-flow workflows. Migration must map model clients, message types, tool loops, session state, Team or GraphFlow behavior, checkpoints, and telemetry. Run old and new versions against the same task and failure corpus instead of treating the successor label as compatibility proof.
Keep business effects outside framework guarantees
Workflow checkpoints and hydration preserve framework state, but they cannot prove whether an external payment, email, deployment, or database write committed before a crash. Bind approvals to canonical arguments and resource versions, use durable idempotency keys and effect records, reconcile unknown outcomes, and enforce tenant and object authorization at the service boundary.
Key Characteristics
- Stable core: Python and .NET reached version 1.0 with stable agent and workflow APIs
- Preview Go line: the separate Go SDK covers core agents and workflows but documents feature gaps
- Agent composition: combines model clients, instructions, tools, sessions, context providers, and middleware
- Typed workflows: route data through graph edges and executors rather than relying only on free-form conversation
- Operational controls: provide checkpointing, pauses, streaming, telemetry, and long-running workflow support
- Application control: protocol adapters do not replace authorization, isolation, or side-effect consistency
Common Use Cases
- Building a Python or .NET assistant on the stable 1.0 API
- Evaluating the preview Go SDK for bounded agents and workflows
- Composing deterministic functions and model-driven agents in one typed workflow
- Pausing long-running work for authorization and resuming from persisted state
- Migrating AutoGen or Semantic Kernel applications with explicit behavioral mapping
Example
Loading code...Frequently Asked Questions
Is Microsoft Agent Framework production-ready?
The Python and .NET core reached version 1.0 in April 2026 with stable APIs and a long-term-support commitment. Individual integrations can still be preview. The separate Go SDK is also public preview, so teams must record the language, package version, and feature status they actually deploy.
Is Microsoft Agent Framework the replacement for AutoGen?
Microsoft describes Agent Framework as the direct successor to AutoGen and Semantic Kernel for new development, while Microsoft AutoGen is in maintenance mode. Existing applications still require explicit mapping of clients, messages, tools, sessions, state, orchestration, and telemetry.
Does Microsoft Agent Framework support Go?
Yes, through a separate open-source Go repository, but that SDK is in public preview and does not yet match every Python or .NET feature. Evaluate the exact provider, workflow, hosting, declarative-agent, and orchestration capabilities required by the workload.
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. Use a Workflow when routing, parallelism, approvals, or recovery boundaries must be explicit. A Workflow can include agents alongside deterministic executors, so the abstractions are complementary.
Does workflow checkpointing guarantee exactly-once execution?
No. A checkpoint records framework state but cannot prove whether an external effect committed before a crash. Consequential tools need stable operation IDs, idempotency, durable effect records, reconciliation, authorization, and compensation where reversal is possible.