01 · Magazine · Comparison
Agency Swarm vs. CrewAI: which framework fits your agent team?
In short: for most teams, CrewAI is the better choice, with broader model support, Flows for fixed processes, built-in persistence and a much larger community. Agency Swarm pays off if you already build on the OpenAI Agents SDK and want to define exactly which agent may talk to which. Both frameworks are free under the MIT license (as of October 2026).
Multi-Agent Navigator editorial team · Reviewed
AAgency SwarmFor OpenAI-centric teams
Agency Swarm models agent teams like an organisation chart with directed communication flows. Since version 1.0 it builds on the OpenAI Agents SDK and the Responses API; the current version is 1.11. Other providers work via LiteLLM.
View profileBCrewAIFor most team scenarios
CrewAI is a standalone Python framework (not based on LangChain) with role-based crews and event-driven Flows for fixed processes. It supports all major model providers and offers an optional commercial platform. The current version is 1.15.
View profileDetailed comparison
| Criterion | Agency Swarm | CrewAI |
|---|---|---|
| Architecture | Agency with directed communication flows (ceo > dev) | Crews with roles and tasks, plus Flows |
| Foundation | OpenAI Agents SDK | Standalone framework |
| Version (October 2026) | 1.11 | 1.15 |
| Python | 3.12 or later | 3.10 to 3.13 |
| Model support | OpenAI native, others via LiteLLM | All major providers natively and via LiteLLM, Ollama |
| Fixed processes | Via communication flows and your own code | Flows with state, routing and persistence |
| Persistence | Callbacks to store threads in your own database | Flow state, memory and knowledge built in |
| Human review | Via your own tools and code | human_input=True on tasks: the agent asks before its final answer |
| Prompt control | Full control over every agent's instructions | Role, goal and backstory; framework adds its own prompt parts |
| Learning curve | Low, if you know the OpenAI Agents SDK | Low, many tutorials and templates |
| GitHub stars (October 2026) | about 4,500 | about 59,000 |
| License and costs | MIT, framework free; paid Agents-as-a-Service only from the author | MIT; optional platform with free tier (50 executions/month) |
Which one for which situation?
- You are starting your first multi-agent project
- CrewAI. The CLI creates a project with agents and tasks, and you will find an example for almost every question. If you outgrow crews, Flows add fixed processes with state.
- Your product already runs on the OpenAI Agents SDK
- Agency Swarm. It adds directed communication flows and a clear role model on top of the SDK you already use, without introducing a second framework concept.
- You need models from several providers or local models
- CrewAI. It connects the major providers natively and Ollama for local models. Agency Swarm can do this via LiteLLM, but is designed around OpenAI.
- Every message between agents must follow strict rules
- Agency Swarm. You define which agent may send messages to which, and you write every prompt yourself. That makes the communication easy to audit.
- You want monitoring and hosting without your own infrastructure
- CrewAI with its platform. The free plan covers 50 workflow executions per month; Enterprise adds SSO, role-based access and deployment in your own VPC.
Frequently asked questions
Is Agency Swarm or CrewAI better for beginners?
CrewAI, if you do not know any framework yet. There are many tutorials and templates, and a CLI that creates a project with agents and tasks. Agency Swarm is easy if you already know the OpenAI Agents SDK, because it builds directly on it.
Can I use Agency Swarm without OpenAI models?
Yes. OpenAI models are supported natively, while Anthropic, Google Gemini, Azure OpenAI and other providers work through LiteLLM. The framework's structure follows the OpenAI Agents SDK and the Responses API, though, so it is easiest to use in an OpenAI setup.
How much do Agency Swarm and CrewAI cost?
Both frameworks are licensed under MIT and free of charge. You only pay for model usage. CrewAI also offers a platform with a free plan (50 workflow executions per month) and an Enterprise plan on request. The author of Agency Swarm markets a paid Agents-as-a-Service offering; the framework itself has no platform plan of its own.
Which Python version do I need?
Agency Swarm requires Python 3.12 or later. CrewAI runs on Python 3.10 to 3.13. If your project still runs on Python 3.10 or 3.11, you can use CrewAI without upgrading.
How do I store conversation histories permanently?
In Agency Swarm you pass two callbacks to the Agency (load_threads_callback and save_threads_callback) and use them to store histories in your own database. CrewAI comes with memory, knowledge and the persisted state of Flows, so you need to build less yourself.
Choose Agency Swarm if
- you already work with the OpenAI Agents SDK
- you want to define exactly which agent may talk to which
- you want full control over every prompt
- a small, free MIT framework without its own platform is enough
Choose CrewAI if
- you want a large community with many examples
- you need several model providers or local models
- you want fixed processes with state (Flows) next to autonomous crews
- a managed platform with monitoring may be needed later
Our verdict
For most teams CrewAI is the more robust choice: broader model support, Flows for predictable processes, built-in persistence and a much larger community.
Agency Swarm makes sense if you already build on the OpenAI Agents SDK and want explicit, directed communication between agents with full control over prompts.
Author: Multi-Agent Navigator editorial team
How we research and review: About and methodology
Last reviewed:
Whitepaper · Edition 2026-Q4
Multi-agent systems in the enterprise
6 chapters, 17 sources, free PDF.
Get the whitepaper