Microsoft Agent Framework now supports multi-agent orchestration at GA, and the most interesting part is the Magentic pattern — a dynamic orchestration approach that picks which agent handles what based on the task, rather than following a fixed workflow.
The patterns available
The framework ships with five orchestration patterns out of the box. Sequential runs agents one after another. Concurrent runs them in parallel. Handoff passes control from one agent to the next. Group chat lets agents communicate freely. And Magentic, which is the one worth paying attention to, dynamically selects and delegates to agents based on the current task state.
Most frameworks give you the first four. Magentic is the differentiator. Instead of a hardcoded chain of agents, the orchestrator decides at each step which agent is best suited for the next action. This means the system can adapt to unexpected situations without you having to predefine every possible path.
What Magentic actually does
Magentic is a dynamic orchestration pattern where a lead agent (the orchestrator) maintains a plan and delegates subtasks to specialized worker agents as needed. The orchestrator can create new agents, reassign tasks, and revise the plan based on intermediate results. This is different from fixed patterns where agent A always passes to agent B, which always passes to agent C.
The approach was validated against Magentic-One, Microsoft’s research system, which scored 38 percent on the GAIA benchmark. That benchmark tests general AI assistant capability across a range of tasks including web research, document analysis, and multi-step problem solving. The same system also showed competitive results on AssistantBench and WebArena, which test web-based interaction and navigation.
To be clear, 38 percent on GAIA is not a headline number. The benchmark is hard, designed to be challenging for current systems. The more relevant point is that a dynamic orchestration approach can match or exceed fixed patterns on structured tasks while being significantly more flexible on open-ended ones.
Why this matters for agent builders
If you are building multi-agent systems today, you are probably using fixed orchestration patterns. Agent A researches, Agent B writes, Agent C reviews. This works for predictable workflows but falls apart when the task diverges from the expected path. Dynamic orchestration handles the divergence by letting the system decide the next step based on actual results, not predefined rules.
The practical benefit is simpler agent definitions. You do not need to anticipate every possible path. You define the available agents and their capabilities, and the Magentic orchestrator figures out the routing. This reduces the amount of orchestration logic you need to write and maintain.
Practical considerations
Dynamic orchestration comes with tradeoffs. The orchestrator agent itself makes the routing decisions, which means it needs to be a capable model. Running a dynamic orchestrator on a weak model will produce poor routing decisions, which defeats the purpose. You need to budget for the cost of the orchestrator’s reasoning calls on top of each worker agent’s work.
There is also the question of predictability. Fixed patterns are easier to debug because the execution path is deterministic. Dynamic patterns can produce different outcomes for the same input depending on the orchestrator’s decisions. This is acceptable for exploration and research tasks but may be a problem for compliance-sensitive workflows.
For most teams, the right approach is to start with fixed patterns for well-understood tasks and use Magentic for the parts of the workflow that need flexibility. The framework supports mixing patterns within the same system, so you can use sequential for the stable parts and dynamic for the exploratory parts.
Getting started
The patterns are documented in the Agent Framework documentation with sample code for both .NET and Python. A good place to start is the group chat pattern for simple multi-agent coordination, then graduate to Magentic once you have a handle on how your agents interact. The GitHub repo has examples for each pattern.
Comparing Magentic with other multi-agent frameworks
Magentic is not the only dynamic orchestration framework out there. LangGraph from LangChain offers a graph-based approach where you define nodes and edges, and the system navigates the graph dynamically based on state. CrewAI uses role-based agent definitions with task delegation. AutoGen from Microsoft Research was an earlier research project that informed parts of the Agent Framework.
The main difference is that Magentic is built into the Agent Framework as a first-class pattern rather than a separate framework. You get it alongside the other four patterns, the Copilot and Claude Code connectors, and the built-in tool surface. This means less time gluing frameworks together and more time building agents.
For teams already using LangGraph or CrewAI, there is no immediate reason to switch unless you want the deeper Azure integration or the built-in connectors. For teams starting fresh, the Agent Framework with Magentic is worth evaluating because it gives you multiple orchestration patterns in a single package instead of requiring you to pick one framework upfront.
What to watch for
Multi-agent orchestration is still an emerging area. The patterns and best practices are evolving rapidly. Magentic in its GA form is a solid foundation, but expect the patterns to get more sophisticated over the next year as Microsoft and the community build on top of them. The framework’s pattern-based design means new orchestration strategies can be added without breaking existing agents, which is the right architectural choice for a fast-moving field.