Building a Digital Workforce with Multi-Agents in Azure AI Foundry Agent Service

As organizations increasingly rely on AI to automate complex tasks and scale digital operations, the ability to coordinate multiple agents in a single, cohesive system is becoming critical. Moving beyond single-agent architectures to multi-agent systems enables richer, more dynamic automation – where specialized agents can collaborate, share context, and complete multi-step processes with minimal human intervention. This shift is unlocking the potential for organizations to build full digital workforces that can manage everything from customer support to supply chain automation.

 

Why Multi-Agents Matter

Building a single agent is often straightforward – it’s designed to perform a specific task, like answering common support queries or generating summaries from documents. However, real-world processes are rarely this simple. They often require multiple steps, context switching, and complex decision-making, which single agents struggle to handle alone. Multi-agent systems solve this by distributing specialized tasks across multiple agents, each optimized for a specific function, while maintaining coordination and context throughout the workflow.

For organizations, this means:

• Scalability – Tasks can be distributed among multiple agents, which enables the system to scale horizontally when workloads grow (e.g., more users, more data). More agents can handle more tasks simultaneously without bottlenecks.
• Specialization – Agents can be fine-tuned for modular tasks or domains (e.g. research, summarization), improving overall performance and making the system easier to build, test and maintain.
• Flexibility – Agents can be reused across different workflows or composed into new systems. Workflows are easily extended as business requirements evolve and new agents are added.

Introducing Multi-Agent Capabilities in Azure AI Foundry Agent Service

With the public preview of multi-agent capabilities Azure AI Foundry Agent Service, we are announcing several new multi-agent capabilities that provide developers with powerful, yet easy to integrate tools to build, orchestrate, and scale multi-agent systems.

1. Connected Agents – Building Collaborative, Task-Specific Systems easily (Public Preview)

Connected Agents allow developers the simplest way to break down complex tasks into coordinated, specialized roles—without the need for a custom orchestrator or hand-coded routing logic. With this capability, you can design systems where a primary agent intelligently delegates to purpose-built sub-agents, streamlining workflows like customer support, market research, legal summarization, and financial analysis.

Rather than overloading one agent with too many skills, you can build focused, reusable agents that collaborate seamlessly—scaling both performance and maintainability.

This model is particularly useful for scenarios where agents need to perform discrete tasks independently, like data extraction, risk assessment, or personalized content generation.

Key Features:

• Simplified workflow design: Break down complex tasks across specialized agents to reduce complexity and improve clarity.
• No custom orchestration required: The main agent uses natural language to route tasks, eliminating the need for hardcoded logic.
• Easy extensibility: Add new connected agents (for example, translation or risk scoring) without modifying the main agent.
• Improved reliability and traceability: Assign focused responsibilities to each agent for easier debugging and better auditability.
• Flexible setup options: Configure agents using a no-code interface in the Foundry portal or programmatically via the Python SDK.

Connected Agents can be used to automate and reimagine a wide variety of business processes. Consider a Sales Assist Agent that helps sales teams prepare for customer meetings. This system might include specialized sub-agents for tasks like Market Research, Competitive Analysis, Customer Insights, and Financial Analysis. Each sub-agent can independently gather, process, and summarize data, while the main agent compiles and delivers the final briefing, reducing manual effort and improving response times.

Early preview customers have been experimenting with Connected Agents to automate their business processes with multi-agent systems with a lot more velocity than using customer orchestrators.

“Connected agents provide us with a faster, more reliable, and robust way to bring AI-driven solutions into production, aligning perfectly with our mission to seamlessly deliver enterprise-ready AI solutions at speed and scale.” — Mallesh Dasari, Senior Director, Digital Solutions Architecture, NTT DATA Generative AI Technology Hub

 

2. Multi-Agent Workflows – Orchestrating Complex, Long-Running Processes (Coming Soon)

Developers requiring fine grained control over planning, conversation, state management can compose a multi-agent workflow. Workflows introduce a structured, stateful orchestration layer that coordinates multiple agents across multi-step processes. Unlike connected agents, these workflows can handle context sharing, persistent state management, and error recovery over long durations – making them ideal for enterprise scenarios like customer onboarding, financial transaction processing, and supply chain automation.

Example of a multi-agent workflow in action

Workflow Features:

• State Management – Define states to organize agent interactions, grouping agents into logical units based on their roles in the workflow.
• Flexible Transitions – Use rule-based or LLM-driven transitions to define the logical flow between states.
• Structured Data Handling – Use variables to pass structured data between agents without risk of overwrites, ensuring data integrity.
• Durability and Resilience – Built-in persistence and failure recovery mechanisms ensure workflows can handle long-running processes without data loss.
• Visual Design and Debugging – Integration with VS Code for visual workflow design, real-time debugging, and monitoring (coming soon).

 

Data Flow Management
Managing data flow between agents is a critical part of multi-agent program design. Workflows support a wide range of data exchange patterns, allowing agents to share information either by reference (e.g., thread history) or by value (e.g., structured responses). Key data types supported include:

• Messages – Multi-modal outputs generated by LLMs, including text, images, and other media.
• Structured Responses – Outputs generated by tools or LLMs with user-defined schemas, supporting structured, predictable data formats.
• Threads – Full conversation histories, preserving context across multiple interactions.

Once bound to agents, this data can be applied in several ways:

• Thread Message – Messages are appended to an agent's conversation history before activating the turn, preserving context.
• Prompt Substitution – Meta prompts within an agent can dynamically substitute variables during execution, allowing for flexible, context-aware responses.
• Tool Parameter Substitution – Inputs can be bound to specific tool parameters, enabling more precise control over tool execution.

Control Flow 

Workflows also provide rich control flow capabilities, allowing developers to program the transfer of control between agents to achieve complex goals. This is achieved through States, Transitions, and Triggers, which collectively define the logical progression of a workflow:

• States – Logical execution checkpoints within a workflow, each potentially containing one or more agents. For example, a customer service workflow might start with Authentication, proceed to Triage, and end with Billing Support. If a state has multiple agents, their executions run in parallel, with the state transition only occurring once all agents have completed their tasks.
• Transitions – Define the logical flow between states, either through rule-based triggers (e.g., "if customer authenticated, then transition to billing") or LLM-driven triggers that leverage model reasoning to determine the next step.
• Triggers – Events that initiate state transitions, based on either the outputs of LLM reasoning or the results of tool calls. These can be explicitly defined to respond to specific conditions or inputs.

 

Example Use Case - Customer Billing Support: Consider a customer service scenario at a telecom company where a user inquires about their latest bill. The first agent authenticates the customer by verifying identity through email and the last four digits of their SSN. Once verified, the workflow transitions to a billing agent that retrieves account-specific information such as outstanding balances, due dates, or billing history. This modular agent design enables secure handling of sensitive data, clean separation of concerns, and easy extensibility, for example, adding a payment resolution or escalation agent. With long-term state management and human-in-the-loop controls, this multi-agent setup is more robust and maintainable than a single-agent solution handling everything at once.

Customers worldwide are leveraging multi-agent workflows in Foundry Agent Service to transform complex business processes, from accelerating software development lifecycles to simplifying enterprise compliance and testing:

"At JM Family, we leveraged Azure Foundry AI Agent Service to develop BAQA Genie - a collaborative ecosystem of AI agents that streamlines the software development lifecycle. Each agent specializes in a key phase, from generating user stories to designing test cases, all while keeping humans in the loop to ensure accuracy and alignment. This multi-agent architecture will enable consistent standards, faster turnaround, and measurable improvements in both business analysis and QA productivity across our enterprise."- Amit Sethi, Principal, AI/ML Research Scientist, JM Family Enterprises

 

"Azure AI Foundry Agent Service introduces a powerful and intuitive approach to modeling multi-agent workflows, closely aligning with modern architectures. Its declarative workflow definitions, seamless Azure integration, and developer-focused tooling offer real promise for accelerating our AI initiatives while simplifying management and compliance. We look forward to continued collaboration as its capabilities evolve to meet the needs of enterprise-scale, production-ready AI solutions." – George Tsolis, Distinguished Engineer, Citrix

 

“By partnering with Microsoft to leverage its new agent management capabilities within Azure AI Foundry, NTT DATA can build, manage, and orchestrate AI models across multiple platforms, streamlining complex multi-agent deployments.” - Charlie Doubek, Global VP, Managed Collaboration and Communications, NTT DATA

 

“Renewals are one of the most complex and critical post-sales functions for B2B SaaS companies. They involve coordination across multiple people and systems over months. Azure Foundry AI Agent Service enabled us to build an autonomous system for managing renewals as a set of coordinated, intelligent agents—each with defined goals, guardrails, and handoffs, with built-in persistence and error recovery. The framework strikes the right balance between goal-seeking agent behavior and the predictability and control enterprises need for production systems.” – Prem Parameswaran, Chief Technology Officer, Gainsight

 

3. MCP and A2A Support – Connecting External Agents and Tools

As the multi-agent landscape continues to evolve, open protocols like Model Context Protocol (MCP) and Agent2Agent (A2A) are becoming essential for building flexible, interoperable agent systems. These standards are crucial as they enable developers to seamlessly integrate agents across different platforms, fostering collaboration and task delegation without proprietary lock-in. Azure AI Foundry is committed to embracing these open protocols to ensure developers have the flexibility to build modular, cross-platform agent ecosystems.

We’re excited to announce that Azure AI Foundry supports Model Context Protocol (MCP) and Agent2Agent (A2A) interactions, enabling seamless interoperability across agents and third-party tools. This includes:

1. A2A in Foundry Agent Service:
   Azure AI Foundry Agent Service introduces a new A2A API head, enabling open-source orchestrators to connect with Foundry Agent Service agents without requiring custom integrations. This API head supports multi-turn conversations, seamless context handoffs, and bi-directional communication, making it easier for developers to extend their existing agent systems without rebuilding core logic.

• • Open A2A API Head: Allows third-party orchestrators to invoke agents from Foundry Agent Service, facilitating task delegation and context-aware processing.
  • Multi-Turn Interactions: Enables agents to handle multi-step conversations and pass context fluidly between agents, ensuring consistent responses.
  • Cross-Platform Flexibility: Designed to work with a wide range of open-source agent frameworks, including AutoGen, LangChain, and Semantic Kernel, providing maximum flexibility for developers.

2. MCP and A2A in Semantic Kernel:
   Semantic Kernel provides a powerful layer for integrating MCP and A2A support directly within your agent code, making it possible to connect agents, tools, and data sources through flexible, API-driven interactions.

• • MCP Support: Enables developers to define and consume OpenAPI-based tools within their agent workflows, providing a standardized way to connect with third-party services and APIs.
  • A2A Orchestration: Supports dynamic, context-aware agent composition, allowing developers to create complex, multi-agent workflows with fine-grained control over task delegation and message passing.
  • Unified API Surface: Provides a single, consistent API for integrating external tools, services, and agents, making it easier to build modular, interoperable agent systems at scale.

 

4. Agent Catalog – Accelerating Agent Development with Reusable Code Samples (Public Preview)

The Agent Catalog jumpstarts single and multi-agent system development with a growing library of pre-built, reusable code samples, from Microsoft and trusted partners. Designed for developers and solution architects, the catalog showcases real-world agent configurations and workflows across industries, from customer service to manufacturing optimization. Each sample links to a GitHub repository, where you can explore the source code, extend functionality, and adapt the agents to your unique business needs.

 

 

Trusted Microsoft partners including Auquan, Marquee Insights, MiHCM, Saifr, and Sight Machine are already contributing agent code samples to the catalog, open-sourcing production-ready solutions that address industry-specific challenges. Here’s how some of our partners envision helping developers drive real business impact:

 

"We’re excited to work with Microsoft AI to create agentic solutions that meet clients where they work and solve real problems. In regulated industries, compliance reviews of content are critical but tedious—and as the volume of content increases exponentially, there are human limitations to ensuring that problematic content is caught and corrected before publication. The Saifr Communication Compliance Agent (Saifr CCA) is a compliance guardrail that boosts efficiency in content creation and helps reduce regulatory risk by detecting and correcting potentially risky content. As part of this collaboration with Microsoft, Saifr will open source the agent template as an industry standard." - Vall Herard, CEO, Saifr incubated in Fidelity Labs

 

“In food and beverage manufacturing, the filler is one of the most complex machines and often a bottleneck in the process,” said Kurt DeMaagd, Sight Machine’s Chief AI Officer. “Our Filler Optimization Agent uses AI to identify improvements in settings, understand stoppage causes, and prevent microstops—unlocking hidden performance gains for the entire line. Built on Azure AI Foundry Agent Service, this agent orchestrates data and machine learning models to deliver real-time insights and predictive analytics through a natural language interface, without requiring data science expertise. Our Microsoft collaboration ensures scalability, security and reliability to transform bottling operations into improved throughput.” 

 

“We created the AI News Agent to give decision-makers a clear edge in the rapidly evolving AI landscape, cutting through the noise to surface what truly matters. Built on Azure AI Foundry Agent Service, this solution can be deployed in minutes, delivering curated, high-impact updates tailored to sectors like Microsoft, healthcare, and legal. This code sample proves that lightweight agents can drive serious business value, especially when anchored by enterprise-grade infrastructure.”- Treb Gatte, CEO, Marquee Insights

 

Looking Ahead

As we continue to evolve multi-agent capabilities and Agent Catalog, and embrace new protocols in Foundry Agent Service, expect even tighter integration with Azure’s broader ecosystem, expanded catalog options, and more advanced orchestration features. Stay tuned for more updates and be sure to share your feedback as we shape the future of multi-agent AI together.

 

What’s Next?

• Explore the documentation to learn more about Azure AI Foundry Agent Service
• Start building your agents today in Azure AI Foundry
• Visit our learning modules to build your first agent with Azure AI Foundry Agent Service
• Watch our multi-agent breakout session at Build