controlflow

A Python framework for building agentic AI workflows with task-centric design.

ControlFlow

The dream is to have autonomous AI agents doing everything for us. We will just sit back, relax, and AI agents would know what needs to be done. That is where tools like control flow give us a lot of hope that this might be possible someday. Not at the moment, not in the near future, in my humble opinion, but soon enough. 

And I'm not even talking about artificial general intelligence, where the robots with the help of general AI will be mimicking humans. I'm just simply talking about AI agents or agentic software doing human-level intelligence work, especially when it comes to creativity. 

Control flow is a Python framework for building agentic AI workflows. Control flow provides a structured developer-focused framework for defining workflows and then assigning the work to LLMs without sacrificing control or transparency. There are three features of this new project which is still...

Evolving very very early days. First, it creates discrete observable task for an AI to solve. It is totally task oriented. Then, it assigns one or more specialized AI agents to each task. And then, it combines task into flow to orchestrate more complex behaviors. This task Centric approach allows you to harness the power of AI for complex workflows while maintaining fine grain control. By defining clear objectives and constraints for each task, you can balance AI autonomy with precise oversight. Letting you build sophisticated AI powered applications with confidence. 

In this video, we are going to install this control FL locally and then we will see how it works. Before I do that, let me give a huge shout out to M compute for sponsoring the VM and GPU for this video. If you are looking to rent a GPU on affordable prices, I will drop the link to their website in video description.

Plus I'm also going to give you a coupon code of 50% discount on range of a GPU. So do check them out. Let me take you to my terminal. 

So as you can see, I am running 2 22.041 GB of vram. Next up, let's create a virtual environment with K. I'm just calling it flow. Let's wait for it to get activated and my K environment is ready and we have stepped into it. 

Next up, let's install control Flow by simply doing P install control flow and it is going to install a lot of things. So let's wait for it and control flow is installed. Let me clear the screen and you can see that it has installed a lot of packages mostly around API based models and Lang chain. 

The default model for control flow is open a based API models like GPT 40 and stuff or from Azure AI or Azure openi. But you can use models from anthropic from Gemini and few others by using Lang chain. So you would have to integrate.

Control flow with Lang chin. But as these are very early days, so I think it would be a good idea. If you want to just try it out, go with open. And of course, if you're using openi, you would have to grab the API key from their website, which is platform.

Open.com and that is a paid option. So let me set my key and then clear the screen. So my key is set and I have cleared the screen. 

Next up, let's launch our Jupyter notebook so that we will use this Python library in the Jupyter notebook. Now one question you might ask is that if it supports Ama or not. You might be able to get it working with the help of Lang chain, but it's very rudimentary at the moment. So I would suggest wait a bit and then once this project gets more stabilized in terms of integration, then we can use it. 

I might make another video with Lama integration and my Jupyter notebook is ready. Let me import the libraries which we have installed such as our control flow and Pantic. That is done and now let's create an agent. So I'm just creating a control flow agent with the CF object here. I'm just calling it Deep Thought and the instruction is to

Use a formal tone and clear language and it will be a writing agent. Next up, let's define a class with the search topic with the title and the keyword because you want this agent to research on a topic. And of course, you can change this code as per your own task and your own agent and use case. 

Okay, that is done. Next up, let's define a decorator function. You can see that I'm decorating this research flow research workflow function with this where there is a task one where we are asking that agent to choose a topic. It is going to choose that topic and it is it will be using this research topic class. 

Then it is going to do the online task here. For example, it is going to create the outline based on the topic. After that's done, the task three is that the author agent will write a first draft. 

So all in all, what it is doing is just breaking down a big work into.

Task and then uh tackling those tasks in tandem. So let me run it. 

Finally, all you need to do is to run this research workflow and print the results. And there you go. So this is the outline of the research article on deep thought here where it has gone with the research question, some subtopics, and then it is going about methodology and all that stuff. 

So you can see that the basic idea behind this control flow is that it is designed to address the challenge of building AI powered applications that are both powerful and predictable. What it does is it wholly task centric architecture where it is breaking down complex AI workflows into manageable observable steps. 

Then this approach ensures that the AI agents operate within well-defined boundaries, making it easier to reason about and manage complex workflows. And then it also tries to bridge the gap between AI and

Traditional software with type safe outputs. By using pantic models, you can ensure that AA generated content always matches your applications requirement. And you can build specialized agents which are task specific so that you will have efficient problem solving. 

I believe that they have um become dependent on Lang chain for the models. I think that is a drawback because they should be uh in inad of relying on Lang chain or any other framework. I think they should be relying on themselves because Lang chain has its own uh ecosystem plus agent capabilities. 

So I think this is where they are um duplicating the effort. So instead of going with Lang chain integration, they should offer some out of the Box integration with AMA with the LM studio and then API based model like they are doing uh right now with open a. 

So that's it. Let me know what do you think about this new.

Agentic software. If you like the content, please consider sharing AI Agent Store with colleagues. If you're already subscribed, please share it among your network as it helps a lot. Thanks for watching.

Note, that there may be slight differences in text and video. Watch video for the original content.

ControlFlow Video