Microsoft releases “Microsoft Agent Framework”: open source SDK and runtime, simplifying orchestration of multi-agent systems

Microsoft released Microsoft Proxy Framework (Public Preview) is an open source SDK and runtime that will unify core ideas from Autogen (agent runtime and multi-agent mode) using semantic kernels (enterprise controls, states, plugins) to help teams build, deploy, and observe production output of AI agents and multi-agent worksheets. The framework is available for Python and .NET and is directly connected with Proxy services for Azure AI Foundry For scaling and operation.

What exactly is Microsoft Transport?

• Merged proxy runtime and API surface. The proxy framework comes with Autogen’s single- and multi-agent abstraction, while adding enterprise capabilities to semantic kernels: thread-based state management, type safety, filters, telemetry and extensive model/embedded support. Microsoft positioned it as a successor built by the same team, rather than giving up on a replacement for either project.
• First-class orchestration mode. It supports Agent Orchestration (LLM-driven decision making) and Workflow Orchestration (Deterministic, multi-agent flow of business logic), allowing creative plans to coexist with reliable handover and constraints.
• Pro-code and platform interoperability. base AIAgent The interface is designed to exchange chat model providers and interoperate with Azure AI Foundry agents, OpenAi Assistants, and Copilot Studio, reducing vendor lock-in at the application layer.
• Under the MIT license, open source, multilingual SDK. GitHub Repo publishes Python and .NET packages with examples as well as CI/CD-friendly scaffolding. Autogen is still maintained (bug fixes, security patches) with guidance to consider proxy frameworks for new versions.

Where does it run in production?

Azure AI Foundry Agent Services Provides a managed runtime: it links models, tools and frameworks; manages thread state; executes content security and identity; and wire observability. It also supports Multi-agent orchestration Locally, distinguish yourself from Copilot Studio’s low-code approach by targeting complex Pro Code Enterprise solutions.

But how does it connect to “AI Economics”?

Enterprise AI economics is dominated by token throughput, latency, failure recovery, and observability. Microsoft’s merger provides a run-time abstraction solution for proxy collaboration and tool usage, (b) attaching production controls (TELEMETRY, filters, identity/network, security, security) to the same abstraction, and (c) deploying on managed services that handle scaling, policies, and diagnostics. This reduces the “glue code” that typically drives cost and brittleness in multi-agent systems and is consistent with Azure AI Foundry’s model catalog+ toolchain approach.

Architectural notes and developer surfaces

• Runtime and status: Agents coordinate through the runtime of handling lifecycles, identity, communications and security boundaries – from automatic gene inheritance and form. Thread It is a national unit that can enable reproducible operation, retrieval and audit.
• Features and plugins: The framework relies on the plug-in architecture and function calls of the semantic kernel to bind tools (code interpreters, custom functions) into proxy policies with typing contracts. ((
• Model/Provider Flexibility: The same proxy interface can be targeted at Azure OpenAI, OpenAI, Local Runtimes (e.g. Ollama/Foundry Local) and GitHub models, enabling cost/performance tuning for each task without rewriting the orchestration logic.

Corporate context

Microsoft will publish as part of a broader push to span interoperable, standard-friendly “agent” systems across Azure AI Foundry – with previous statements about multi-agent collaboration, memory, and structured retrieval. It is expected that when stable, closer bonds can be made to observability and governance control.

We like this direction because it crashes two different stacks – Autogen’s multi-agent runtime and semantic kernels – with an API surface with a managed production path. Thread-based state model and OpenTelemetry hooks solve the regular blind spots of proxy systems (Repro, delay tracking, failed classification) and Azure AI Foundry’s proxy services to enable identity security and tool orchestration, so teams can iterate over policies rather than glue code. Python/.NET equality and provider flexibility (Azure OpenAI, OpenAI, Github model, local runtime) can also be practical for cost/percting without rewriting the scheduling.