Introducing Microsoft Agent Framework

On October 1st, Microsoft has announced new capabilities in Azure AI Foundry a simpler way for developers to build, observe, and govern multi-agent systems, while helping organizations strengthen trust and accountability in AI.

As agentic AI adoption accelerates, according to PWC , eight in ten enterprises are now using some form of agent-based AI, and let's not deny that the complexity of managing these systems is growing day after day. Developers are still facing fragmented tools, and organizations might be struggling to guarantee that the agents act with more responsibility. The latest updates from Azure AI Foundry tackle these challenges as they are providing an integrated platform for continuous development, oversight, and governance of agentic AI.

Microsoft Agent Framework - public preview

Now in public preview, the Microsoft Agent Framework is an open-source SDK and runtime that provides the orchestration of multi-agent systems. It gathers AutoGen, which is a former Microsoft Research project with the enterprise-ready foundations of Semantic Kernel, into a single and centralized commercial framework.

What can developers do with the Microsoft Agent Framework?

Today with Microsoft Agent Framework, developers have the possibility to:

implement locally and deploy to Azure AI Foundry with built-in observability, state durability, and compliance
integrate any API via OpenAI with the possibility of collaborating across runtimes with Agent2Agent, and connect to tools with Model Context Protocol dynamically
adopt the modern multi-agent patterns like Magnetic One and orchestrate agents like Workflows
Reduce switching between different contexts across tools and platforms
build multi-agent systems that connect Microsoft 365 Copilot, Azure AI Foundry, and other agent platforms

The main goal of this framework is to keep developers in flow. Based on a recent study, 50% of developers can easily lose more than 10 hours each week to overwhelming tasks dealing with limitations in fragmented tooling that do not guarantee complexity reduction, which leads to a poor developer experience

Multi-agent workflows - private preview :

Microsoft succeeded in bringing major capabilities to the cloud with multi-agent workflows in Foundry Agent Service, since this feature gives the developers the capacity to orchestrate multiple-step business processes through a structured and stateful workflow layer.

With the help of multi-agent workflows, the developers have the capability of:

coordinating many agents on tasks running for a long period with the possibility of sharing the context and persisting the state
automating complex scenarios based on the business enterprise processes
improving reliability at scale with built-in error handling, retries, and recovery

With the current update, workflows can be created visually in the VS Code Extension or Azure AI Foundry, then deployed, tested, and managed in Foundry.

OpenTelemetry powering Multi-agent observability across frameworks :

Microsoft also introduced upgraded multi-agent observability, which includes new OpenTelemetry contributions that aim to standardize tracing and telemetry for agentic systems.

These improvements, which are developed in collaboration with Outshift and Cisco incubation engine, will provide more visibility into agent workflows, tool call invocations, and inter-agent collaboration, which is considered essential for debugging, optimization, and compliance.

With all this effort, Azure AI Foundry now offers unified observability across multiple frameworks that include Microsoft Agent Framework, LangChain, LangGraph, and the OpenAI Agents SDK.

General availability of Voice Live API in Azure AI Foundry

With the last update, the Voice Live API now enables companies to build scalable and production-ready voice AI agents since it offers a unified, real-time speech-to-speech interface that combines speech-to-text, generative AI, text-to-speech (TTS), avatars, and conversational AI in a single and low-latency pipeline.

Responsible AI capabilities - public preview

As agent observability and framework integration advance, it’s important that AI systems operate in a responsible and secure way since they operate on complex enterprise workflows.

Based on the published article of McKinsey 2025 Global AI Trust Survey, the most challenge that van be encountered to adopt AI is the lack of governance and risk management tools. With Responsible AI features, we now have the following:

Task adherence: to keep agents aligned with their assigned tasks
Prompt shields with spotlighting: to build a strong protection against prompt injection and risky behaviors
PII detection: to identify and manage sensitive information