Building Scalable AI Workflows with Microsoft Agent Framework on Azure

Introduction

The Microsoft Agent Framework is an open-source library for building, orchestrating and deploying AI agents within an Azure-native environment. It combines ideas from the Semantic Kernel and AutoGen projects and adds new capabilities for agentic and graph-based workflows.

Designed around principles of modularity, service-based integration, and in-depth observability, the framework enables developers to create scalable, auditable AI solutions with ease.

In this blog, I’ll share how we applied these principles in a recent project where we developed an agentic workflow leveraging Retrieval-Augmented Generation (RAG) to identify and recommend actions for specific types of clauses, and how the framework helped simplify orchestration and governance in a complex AI pipeline.

Core Principle & Architecture

Building on the core principles of modularity, orchestration, observability, and fault isolation, each agent within a workflow should perform a single, well-defined task to enable scalability, fault isolation and auditability, which is critical for developing enterprise-grade AI solutions.

The framework is built around several core services:

• The Agent Registry maintains metadata about all agents in the system, including their responsibilities, endpoints, and orchestration rules to ensure consistent configuration and simple extensibility.
• Service Classes are reusable abstractions for common tasks such as messaging, storage, and model calls. These classes reduce boilerplate code and enforce best practices for integration with Azure services.
• Workflow Orchestration is powered through Azure Service Bus, which enables asynchronous communication between agents through a queueing system for scalability, and this decoupling improves reliability through fault isolation.
• Telemetry and Observability is handled through built-in integration with Application Insights via OpenTelemetry. This captures execution traces, performance metrics, and error logs, allowing full visibility into agent operations.

These services allow for simple integration with Azure services such as Functions, Blob Storage, Cosmos DB, Key Vault and AI Foundry. Traces can be connected to AI Foundry Tracing, where the agent registry can be accessed in a convenient user interface.

Use Case: Contract Analysis

A user would pass a contract through a frontend to be hosted by the multi-agent system.

To illustrate the Agent Framework in action, we built a contract analysis workflow where a user can pass a PDF through a UI and receive an email detailing which clauses within the contract require actioning prior to renewal. For this, we implemented a RAG-based approach where contracts are chunked at the clause level, retrieved by clause type, and analysed by an agent to determine the required action.

The main agents are detailed below:

• Upload Agent: deals with document ingestion and text extraction via Azure Document Intelligence.
• Text Agent: handles semantic chunking, extracting metadata such as section header and page number. The chunks are also sent to an Azure AI Search index.
• Retrieval Agent: Performs hybrid search (vector and keyword) retrieval using a curated query and a list of keywords.
• Clause Agent: Determines if the retrieved chunk matches the clause definition, signals and aliases. An action is then recommended once classification is confirmed.
• Ensemble Agent: Uses an ensemble of LLMs to confirm the clause analysis agent’s findings. This is an important step for quality assurance.
• Output Agents: Separate agents then perform rendering, persisting and sending notifications to users.

When developing, we persisted agent outputs to Blob Storage for simple debugging and observability. We ensured governance through YAML-based prompts and definitions with Jinja templating. A shared library centralised the agent registry, runtime orchestration, client wrappers for consistency across the workflow.

Lessons Learned

As a team familiar with libraries such as LangGraph, we found that the framework significantly improved how we developed and deployed agentic workflows. Integrating Azure services such as AI Foundry was straightforward, and ensuring security through RBAC and Key Vault didn’t require much additional work.

The modular approach allowed developers to concentrate on agents separately though some care was required to ensure correct data flow between agents through Service Bus. Overall, the framework gave us a strong foundation for building a scalable, well-governed agentic solution in the cloud.