Microsoft retires AutoGen and debuts Agent Framework to unify and govern enterprise AI agents

Microsoft’s multi-agent framework, AutoGen, acts as the backbone for many enterprise projects, particularly with the release of AutoGen v0.4 in January. 
However, the company aims to harmonize all of its agent framework offerings and bring more observability capabilities to the forefront as well. Microsoft released the Agent Framework on public preview, which will now essentially be the company’s sole orchestration and agent framework.
Microsoft said AutoGen and Semantic Kernel will “remain in maintenance mode, which means they will not receive new feature investments but will continue to receive bug fixes, security patches and stability updates.”
“For future-facing work, however, the roadmap is centered on Microsoft Agent Framework, and customers should plan migration to capture the benefits of open standards, durability and Azure AI Foundry Integration,” the company said in an email to VentureBeat. 
The company assured existing workloads on AutoGen or Semantic Kernel will be safe because “no breaking changes are planned.”
Microsoft’s move to consolidate agent frameworks into one shows the company’s strategy in agentic AI. By closely tying observability and data protection to the framework for building agents, the company aims to enable the creation of agents through post-deployment, all in one place.
Agent Framework and Foundry
The Agent Framework consolidates AI workloads into a single SDK, combining the capabilities of both Semantic Kernel and AutoGen, allowing users to build AI agents, manage multi-agent deployments, and set up observability systems. 
Sarah Bird, chief product officer for Responsible AI at Microsoft, told VentureBeat in an interview that so many developers and businesses have been rapidly experimenting and adopting AI agents, but needed a way to bring a lot of capabilities together. 
“What’s really exciting about what we’re releasing this week is a lot of capabilities to help people more successfully build and manage agents in a way that, of course, allows them to be powerful,” Bird said. “But, one that also ensures that they are trustworthy by giving you the tools to, you know, observe their behavior and new guardrails to help them stay on task.”
The new framework offers five capabilities for enterprises building AI agents:

Local experimentation before deployment in Azure AI Foundry

API integration through OpenAPI and collaboration across runtimes with A2A and MCP connections

Use Magentic One and other orchestration agents 

Reduce context switching across platforms

Build multi-agent systems across different agent platforms like AI Foundry, M365 Copilot or others

Microsoft is also adding Agent Framework services, such as multi-agent workflows, to its cloud-based Foundry Agent Service. 
Safety, security and monitoring
Bird said one of the differentiators for Agent Framework lies in its responsible AI features. Microsoft added:

Task Adherence, which keeps agents aligned to tasks

PII Detection, which alerts administrators if an agent accesses sensitive data

Prompt Shields that help protect against prompt injection and highlight risky agent behavior. 

“For what enterprises need to think about with agents, I believe, are three important categories,” Bird said. “Number one is the quality of the agent, does it actually work and is it staying and completing the ask. The second is security, both traditional security and new types of risks like prompt injection attacks or leaking sensitive data. And the third thing is management because future organizations will have thousands of agents who could have access to different things and tasks.”
Microsoft will be contributing to the OpenTelemetry standard for observability. Through AI Foundry, developers with agents built on Agent Framework can track quality, performance and cost. AI Foundry does offer OpenTelemetry observability to agents built with other frameworks and not just Agent Framework.
All-in-one agent frameworks
AutoGen competed with other agent builders and multi-agent frameworks from LangChain, CrewAI and LlamaIndex and there’s no doubt Agent Framework would either. 
Microsoft is not the only one hoping to bring all the needed tools to build, deploy and monitor AI agents. LangChain has been building towards offering these tools even as it aims towards a 1.0 release. As agents become more ubiquitous at enterprises, more platforms could look into providing access to building, deployment and observability tools in one.