Microsoft Unveils Agent Framework to Simplify Multi-Agent AI Orchestration

Microsoft has published the Microsoft Agent Framework in public preview, an open-source SDK and runtime for building and operating multi-agent AI systems. The project merges ideas from AutoGen and Semantic Kernel, delivering a single API surface for production-grade agent orchestration across Python and .NET.

Unified SDK and runtime

The Agent Framework carries forward AutoGen’s single- and multi-agent concepts while integrating Semantic Kernel’s enterprise capabilities. That means thread-based state management, stronger type contracts, filters, telemetry hooks, and broad model and embedding support are now available in one framework. Microsoft positions this as a successor built by the same teams, rather than as a replacement that abandons either prior project.

Orchestration modes

The framework supports two primary orchestration styles. Agent orchestration enables LLM-driven decision making and dynamic planning. Workflow orchestration provides deterministic, business-logic multi-agent flows. Together they allow hybrid systems where creative planning from models coexists with reliable handoffs, constraints, and deterministic steps.

Pro-code interoperability and provider flexibility

The base AIAgent interface is designed to be provider-agnostic. Agents can swap chat model providers and interoperate with Azure AI Foundry Agents, OpenAI Assistants, and Copilot Studio. Supported runtimes include Azure OpenAI, OpenAI, local runtimes such as Ollama or Foundry Local, and GitHub Models. This lets teams tune cost and performance per task without rewriting orchestration logic.

Open-source packages and deployment targets

Microsoft published Python and .NET SDKs under the MIT license, with examples and CI/CD-friendly scaffolding in the GitHub repo. AutoGen will remain maintained for bug fixes and security patches, while teams are encouraged to consider Agent Framework for new builds. For production, Azure AI Foundry’s Agent Service offers a managed runtime that links models, tools, and frameworks, manages thread state, enforces content safety and identity, and provides observability.

Enterprise economics and operations

Microsoft ties the consolidation to enterprise AI economics: token throughput, latency, failure recovery, and observability. By giving a single runtime abstraction for agent collaboration and tool use, attaching production controls like telemetry, filters, and identity, and deploying onto a managed service, the framework reduces glue code and operational brittleness. That aligns with Azure AI Foundry’s approach of model cataloging, toolchains, and governance controls.

Developer surface and architecture notes

Runtime and state are modeled around threads as the unit of execution, enabling reproducible runs, retries, and audits. The framework leverages Semantic Kernel’s plugin architecture and function-calling to bind tools and custom functions into agent policies with typed contracts. OpenTelemetry hooks and thread-based state aim to make latency tracing, failure triage, and reproducibility more practical for teams.

Why this matters

By collapsing AutoGen’s runtime and Semantic Kernel’s enterprise plumbing into a single API surface with managed production paths, Microsoft reduces the burden of glue code in multi-agent systems. The result is more manageable orchestration, clearer observability, and practical provider flexibility for cost and performance tuning.