Bring your towel and join one of the fastest growing data communities. Welcome to our second-generation open source orchestration platform, a completely rethought approach to dataflow automation.

Prefect Community

Going through the docs and ControlFlow looks very interesting. I've been using AutoGen to build my agents, how would you compare ControlFlow vs. AutoGen?

Hey <@U07A9C69961> ! I think the primary difference is that ControlFlow is task-centric and AutoGen is agent-centric.

In practice this means that to use ControlFlow you'll build workflows by describing the objectives you have, then delegating that work to your agents. In AutoGen you'll build workflows by getting a team of agents together, then asking them to fulfill your objectives.

The reason we take this approach is that our goal is not to chat with agents, but to have agents complete non-algorithmic parts of a workflow. So in ControlFlow you can assess the completion of any step of the workflow by working with the `Task` object like any other Python object. It has a status and result. In AutoGen, I'm not sure how you get objective representations of workflow progression because I don't think non-conversational workflows is a major goal

Both are designed to support complex multi-agent setups, but I think like all software choosing the one whose central abstraction meets your mental model will give you the best result!

This is an excellent response thank you! Does controlflow use Marvin? I saw it because it was referenced in this video <https://youtu.be/yj-wSRJwrrc?si=H4ntRtOhzATR2O3r&amp;t=360>

I created a Claude Project and a OpenAI custom gpt using your documentation. The claude project works awesome but it won't let me share it publicly, custom gpt doesn't work at all which is a bummer.

Having your documentation repo exportable into LLM friendly format would be a big help

Marvin's agent APIs only support the OpenAI assistants API, so while CF started with Marvin as the engine, we got immediate requests to expand to other LLMs and decided to give CF its own backend to support that rather than adding first to Marvin. (more for velocity reasons than anything else). So CF is _spiritually_ dependent on Marvin but right now separate. Hope to reunite them in the future.

Love that idea -- I have the same Claude project, it didn't occur to me to that others would want the same! There's even an AI Style guide in the docs because of this. If you open an issue for it, I'm sure we can make it easy. For light questions, our docs provider (Mintlify) automatically builds a conversational API into the docs search bar.

Yea it seems like for building agents and quick debugging you really need Sonnet 3.5 so looking for more of the customgpt/projects interface.

Does the OpenAI api key variable support OpenRouter (same api structure)

Also how do I actually run this either locally or deploy it.

Haven’t used open router so I’m not sure - but CF accepts any langchain compatible model so id bet there’s a way to configure it either on the existing OpenAI model or as a OpenRouter specific one 