Microsoft Agent Framework Hits RC Milestone
Microsoft's Agent Framework for orchestrating AI agents has reached Release Candidate status for both .NET and Python, signaling its readiness for production use in enterprise environments. The framework enables developers to build and manage agents that can execute complex workflows and integrate with LLMs. To support the maturing ecosystem, Microsoft has also published migration guides for developers moving from its earlier Semantic Kernel and AutoGen projects.
- The Agent Framework is the designated successor to Microsoft's earlier Semantic Kernel and AutoGen projects, merging their capabilities into a single, unified open-source platform. This consolidation addresses developer feedback about fragmentation, combining AutoGen's multi-agent orchestration with Semantic Kernel's enterprise-grade features like state management and telemetry. All future development will focus on the new framework, with the legacy projects now in maintenance mode. - A key feature of the framework is its use of graph-based workflows, which allows developers to connect agents and functions with explicit control over the execution path. This supports complex orchestration patterns, including sequential, concurrent, and handoff scenarios, and includes capabilities like checkpointing and support for human-in-the-loop interventions. - To enhance the developer experience, the framework includes an interactive DevUI for testing and debugging agent workflows. Observability is another core tenet, with built-in support for OpenTelemetry to monitor agent actions, tool invocations, and performance. - The framework is designed for interoperability and is model-agnostic, supporting connections to a wide array of LLMs from providers like Azure OpenAI, OpenAI, Anthropic, AWS Bedrock, and local models via Ollama. It also embraces open standards like Agent-to-Agent (A2A) and Model Context Protocol (MCP) to facilitate communication between different agents and tools. - For enterprise environments, the framework integrates with Azure AI Foundry for deployment, providing a managed runtime that handles state management, security through Microsoft Entra integration, and responsible AI features like prompt injection protection. - A practical application of the framework has been demonstrated in a contract analysis workflow, where a team of specialized agents collaborates to process a PDF. This system uses a Retrieval Agent for search, a Clause Agent for analysis, an Ensemble Agent for quality assurance, and Output Agents for notifying users, showcasing a multi-agent solution. - The framework provides middleware to intercept and act on agent-to-LLM requests and agent-to-function calls. This allows developers to insert custom logic, such as logging, between an agent and the tools or language models it interacts with.
