Implement generative AI and agentic solutions 2.2 Build agents by using Foundry
A research task needs several specialized agents (a web researcher, a coder, and an…
Study guide skill: Implement orchestrated multi-agent solutions
Question B8-022
A research task needs several specialized agents (a web researcher, a coder, and an analyst), but the steps and their order aren't known in advance and may need several rounds of reasoning. You want a manager that plans, tracks progress, and picks the next agent as the work evolves. Which Microsoft Agent Framework orchestration fits?
Select one answer.
Answer: B
Magentic orchestration uses a manager that coordinates specialized agents, tracks progress, adapts in real time, and fits tasks where the solution path isn't known in advance. Group chat has the same architecture without complex planning and is the choice when simpler coordination is enough. Sequential runs agents in a fixed pipeline, and concurrent runs them in parallel.
Deep dive
How it works
Magentic orchestration is based on the Magentic-One system from AutoGen and is built for complex, open-ended tasks. A dedicated Magentic manager coordinates a team of specialized agents. It decides which agent acts next based on the evolving context, how far the task has progressed, and what each agent can do. The manager keeps a shared context, tracks progress and adapts the workflow in real time, so it suits work where the solution path isn't known in advance and may take several rounds of reasoning, research and computation.
The workflow builder exposes the inner-loop limits: maximum coordination rounds, maximum consecutive stalls before replanning, and maximum plan resets. It also has a flag for human-in-the-loop plan review. While the workflow runs, Magentic emits events when the initial plan is created and when a new plan replaces it, either after stall detection or because a human revised it. Each coordination round also produces a progress ledger update, which records whether the request is satisfied, whether progress is being made, and who speaks next.
Keep two caveats in mind. You can define your own specialized agents, but the page says it's untested how well Magentic performs outside the original Magentic-One design. Also, the .NET samples mark the Magentic types as experimental in code (pragma MAAIW001).
The other options, and when you'd use them
- Group chat with a round-robin manager: An orchestrator picks who speaks next in a shared conversation, using round-robin, prompt-based selection or custom logic. It's good for iterative refinement and multi-perspective analysis. Magentic has the same architecture plus a manager that plans, so choose group chat when you need simpler coordination without complex planning.
- Sequential: Agents form a pipeline, and each one passes its output to the next. It's ideal when each step builds on the previous one, as in document review, data processing pipelines or multi-stage reasoning. The order has to be known up front.
- Concurrent: Agents work on the same task in parallel and independently, and their results are collected and aggregated. Use it for brainstorming, ensemble reasoning or voting.
Exam tip
"Unknown path, and a manager that plans and tracks progress" is Magentic. "Fixed order" is sequential, "parallel and independent" is concurrent, "shared conversation with simple speaker selection" is group chat, and "agents transfer control to each other" is handoff.
References
Last reviewed: