November 18, 2026

CrewAI vs Microsoft Agent Framework: crews or workflows

Reviewed Nov 2026

CrewAI is a Python framework for crews that run under a sequential or hierarchical process. Microsoft Agent Framework is an open-source SDK built for .NET and Python that pairs agents with graph-based workflows. Pick CrewAI if your team writes Python and wants a fixed task order or a manager-led crew. Pick Microsoft Agent Framework if you want handoff or group chat patterns, or plan to host on Azure Functions with C# or Python.

At a glance

Python agent crewsAgents and workflows SDK
What it isCrewAI sets a process on each crew, and its docs say processes orchestrate the execution of tasks by agents, much like project management in human teams.Microsoft Agent Framework is the direct successor to Semantic Kernel and AutoGen, created by the same teams.
LanguagesCrewAI is tagged as a Python framework in its documentation keywords.Microsoft Agent Framework currently supports .NET (C#) and Python. Its Go version is in public preview and requires Go 1.26 or later.
How agents are coordinatedCrewAI offers a sequential process and a hierarchical one, chosen per crew.Microsoft Agent Framework provides five built-in orchestration patterns: sequential, concurrent, handoff, group chat and Magentic.
State and checkpointingCrewAI checkpointing lets a crew, flow, or agent resume after a failure, and two storage providers ship with it, one for JSON files and one for SQLite.Microsoft Agent Framework creates checkpoints at the end of each superstep and ships three built-in storage implementations for in-process memory, local disk and Azure Cosmos DB.
Human-in-the-loopCrewAI offers two main approaches, flow-based and webhook-based, and its docs label the webhook approach Enterprise.Microsoft Agent Framework orchestrations support human input through tool approval and request info.
ProtocolsCrewAI's event list includes MCP connection events and A2A delegation events.Microsoft Agent Framework includes MCP clients for tool integration, and its Python Foundry chat client has a tool that exposes a remote A2A agent to a Foundry agent.
LicenseCrewAI's license file grants permission free of charge and lists the rights to use, copy, modify, merge, publish, distribute, sublicense and sell copies of the software.Microsoft Agent Framework is open source and available on GitHub.

How does each one coordinate agents?

CrewAI fixes the coordination strategy when you create the crew, by specifying the process type. In the sequential process, tasks follow the order of the task list, and the output of one task serves as context for the next. The hierarchical process needs a manager_llm or a custom manager_agent, which handles creating and managing tasks for the crew.

Workflows connect agents and functions through explicit execution paths. In a handoff orchestration, each agent can pass the entire task to another and no central authority manages the workflow. Group chat places an orchestrator in the middle of a star topology, and that orchestrator can pick the next speaker by round-robin, prompt-based selection, or custom logic.

Magentic and CrewAI's hierarchical process both put a manager in charge. In Magentic, the manager dynamically selects which agent should act next based on the evolving context. Plan review lets a person review and approve the manager's proposed plan before execution, and it is off by default in Python and on by default in .NET.

When should you choose CrewAI?

  • A Python team that wants either a fixed task order or a manager-led crew gets both from the sequential and hierarchical processes.
  • Failed runs need to pick up where they stopped, and a CrewAI run can also be forked into an alternate branch.
  • Checkpoint files you can open and read help your debugging, since the JSON provider writes one file per checkpoint that the docs call human-readable.
  • Reviews happen at a console during local development, which the flow-based @human_feedback decorator is listed as best for, from version 1.8.0.

When should you choose Microsoft Agent Framework?

  • Your services are written in C#, and Agent Framework supports .NET (C#) alongside Python.
  • Handoff, group chat, concurrent or Magentic coordination is what you need, and the framework provides several built-in multi-agent orchestration patterns for them.
  • Checkpoints must survive across processes and machines, which fits the Azure Cosmos DB storage provider listed for production, distributed, cross-process workflows.
  • Azure Functions is your target, and the Durable Extension offers it as managed, serverless hosting in C# and Python, with durable agents that can pause for human input without consuming compute or model tokens while waiting.
  • Some tool calls need sign-off, and agents can use approval-required tools that pause the workflow for human review.

Can you use them together?

Treat any link between a crew and an Agent Framework workflow as integration code you write and test. In a shared system, a CrewAI crew would own its tasks and process, while a Microsoft Agent Framework workflow would own its executors, agents and checkpoints.

A2A appears in both sets of docs, so it is one boundary worth testing. CrewAI lists A2A delegation events, and the Python Foundry chat client in Microsoft Agent Framework has a tool that exposes a remote A2A agent as a tool. Confirm the path works with your versions before relying on it. For background on the two protocols, read A2A vs MCP framework checkboxes, and for other pairings see LangGraph vs CrewAI and LangGraph vs Microsoft Agent Framework.

What should you watch out for?

  • CrewAI's automatic, event-driven checkpoint writes are described as best-effort, and callbacks are saved only as references to named functions.
  • The webhook-based human review path in CrewAI is labeled Enterprise, so check your plan if you want reviews routed through Slack or Teams.
  • Sequential, concurrent and group chat orchestrations in Microsoft Agent Framework do not pause for free-form input on their own, and the docs pair them with a RequestPort in a custom workflow. Microsoft also notes it is untested how well Magentic orchestration will perform outside the original Magentic-One design.
  • The Getting started section of Microsoft's Durable Extension page says to keep the prerelease options shown for preview packages. If your threat model does not permit pickle-based serialization, the checkpoint docs say to use InMemoryCheckpointStorage or a custom storage with another serialization strategy.

Does Microsoft Agent Framework only run on Azure?

No. The Durable Extension also supports bring-your-own-compute hosting in your own worker process, container or Kubernetes environment. The overview lists Anthropic, OpenAI and Ollama among the supported model providers.

Does the hierarchical process in CrewAI need a manager?

Yes. The docs say a manager_llm or manager_agent must be specified in the crew to enable it.

When should a task be a workflow instead of an agent?

Microsoft's overview suggests a workflow when the process has well-defined steps or when multiple agents or functions must coordinate. It also says that if you can write a function to handle the task, you should do that instead of using an AI agent.

Can AutoGen users move to Microsoft Agent Framework?

Microsoft's overview points to a Migration Guide from AutoGen, and the Magentic pattern is based on the Magentic-One system invented by AutoGen. See Microsoft Agent Framework vs AutoGen for that pairing.