Microsoft's Agent Framework Details Five Orchestration Patterns for Multi-Agent AI

Microsoft's Agent Framework documents five orchestration patterns for running multiple AI agents together, each with sample code and a use case. The patterns—concurrent, sequential, group chat, handoff, and Magentic—are built around the same Foundry chat client and require an Azure AI Foundry project with a deployed model such as gpt-4.1-mini.

Concurrent orchestration is the simplest pattern. Every participant receives the same input and responds independently; the workflow collects the results. The sample routes a customer support ticket to sentiment, category, and priority agents at the same time, which suits situations where multiple perspectives or votes are needed.

Sequential orchestration chains agents so each step consumes the previous agent’s output. The sample condenses a ticket into a one-sentence summary, then passes that summary to a classifier that assigns exactly one category. This pattern fits tasks that decompose into dependent stages.

The group chat pattern uses a central manager agent to choose which participant speaks next and can pause to ask a human for input. In the example, product, engineering, design, and security agents debate a feature proposal while a manager moderates and eventually issues a recommendation. The workflow has a max_rounds parameter, so the discussion ends after a fixed number of turns even if the participants would otherwise continue.

The handoff pattern differs: no manager is in charge. Agents decide on their own to transfer the conversation to another participant based on expertise. The sample starts a triage agent that greets a customer, identifies the issue, and routes to a refund or order-status agent; those specialists can hand back to triage if the request changes. This pattern suits workflows where the order of steps cannot be known in advance.

The most elaborate pattern is Magentic, a name used in Microsoft’s training material. A manager agent maintains two stores: a task ledger for remaining work and a progress ledger for completed steps and lessons learned. It assigns agents, updates the ledgers, replans when stuck, and synthesizes a final answer. The sample runs an autonomous blog-writing loop: a writer produces a draft, an editor critiques it, and the manager routes revisions until the piece is accepted. Parameters such as max_stall_count=2 and max_round_count=10 keep the loop from running indefinitely.

The distinguishing axis across the five patterns is control: group chat centralizes turn selection, handoff spreads routing across agents, and Magentic adds planning plus persistence of task state. Microsoft’s samples, available on GitHub, make the patterns easy to try, assuming you already have an Azure Foundry deployment and a model endpoint. The Magentic pattern is covered separately in Microsoft’s training module, which provides additional context beyond the sample code.