I’ve been a “dabbler” with crewAI for a while now, having come across it in March of 2024 and tinkering when I have some time. I love the idea of task-based LLM actions that are specific enough that you can use lower cost but specifically-trained models for the tasks, even running those models on your own hardware. A few weeks back, my team at work used CrewAI for a hackathon in an effort to automate an onerous process, and it came out pretty well!

So, when I saw that they have a couple official training videos on a new e-learning platform called DeepLearning.ai, I figured I’d check them out. #CrewAI is evolving rapidly, and the some of the notes I’ve taken over the past 8 months aren’t even applicable anymore, so I figured this was a great way to level-set and fill in gaps in my knowledge.

I’m not going to describe CrewAI here, other than it’s a framework for easily building multi-agent teams and requires little to no coding experience. How CrewAI works is well-explained in the first fifteen minutes of the course, so at least listen to that part!

About the Course

The course, Multi AI Agent Systems with crewAI, follows a flow familiar to anyone that has taken online courses, and is taught by the creator of crewAI, João Moura. The lessons, ranging from a minute to 18 minutes, are a mix of descriptive lecture and hands-on coding. For the lessons where coding is involved, the window handily split-screens and on one side is an iPython notebook environment with the code pre-populated, so you can work through it as João explains what the code does.

You can also get an indication of the course and CrewAI by checking out my crewAI experimentation repo on GitHub.

Target Audience Analysis

Professionals working in project management, artificial intelligence, and team leadership can greatly enhance their skills in constructing multi-agent teams. Those keen on optimizing team performance, utilizing state-of-the-art technologies for collaborative work, and streamlining task execution processes would discover value in enrolling in specialized online classes tailored to augment their proficiency in this realm. Addressing prevalent challenges like steering diverse teams, accomplishing project goals in intricate scenarios, and keeping pace with evolving team dynamics is indispensable for professionals aiming to excel in their respective roles.

Content Outline

Introduction to Multi-Agent Teams

Familiarizing with the basics of multi-agent teams and their significance in managing complex tasks effectively.

Importance of Building Multi-Agent Teams for Complex Tasks

Unveiling the reasons why seamless collaboration among agents is crucial for successful task fulfillment.

Strategies for Creating Effective Multi-Agent Teams

Scrutinizing established methodologies for assembling and overseeing high-performing multi-agent teams.

Multi-Agent Crew Performance Examples

The bulk of the course is working through five clear examples of building multi-agent systems with CrewAI. The result is a set of real-world instances where multi-agent teams can perform and achieve remarkable results across diverse problems.

write_article

The first example is super simple: take a task you might do with a single prompt with an #LLM chatbot, such as chatGPT, and have it performed by multple agents, each with their own persona. This performs no research and the output is purely from the LLM, making it an easy entrypoint for anyone. (Hint: I had it write an article about taking an online course for developing multi-agent teams, and even incorporated a bit of it into this.)     – this one required an update from the training to run on the latest version of crewai.     – older versions of crewai used an integer for the verbose declaration, and if you are running the latest, you need to change that to a binary, e.g. True.

This example is definitely one you’ll want to revisit after you learn how to use and assign tools to your agents in the following lessons.

customer_support

This creates a customer support agent that can answer questions about a specific product or service. It does this by accessing the URL with the support documentation.

customer_outreach

This example creates some marketing lead material to reach out to a specific company that is a potential customer or partner for a second, i.e. “your” company.

event_planning

This example uses multiple agents to research, identify, create the logistics, and then create some marketing material for an event. It takes in parameters like city, size, and budget in order to find a viable venue. – I believe it was this one where I had to fiddle with the asynchronicity of the agents, since I understand that CrewAI needs to have the last agent to perform a task be performing that itself. I could have that wrong, but I had to change that to make mine work.

This is actually a super-cool example, but I found that the LLMs did not adhere to the parameters, often getting venues too small or ignoring the input I would provide while they were performing their tasks. That’s to be expected, however, and I think experimentation is the name of the game when it comes to building these systems.

resume_builder

The final one was to have the agents create bespoke resumes, based upon the job one is applying for. As opposed to the event planning exercise, the output on this one was very good, and I was impressed with how well it could craft a resume for the specific job, as well as anticipate some of the interview questions and provide some hints for how to answer them.

Conclusion

This course provides a clear and thorough introduction to crewAI, bringing the attendees to an intermediate level of being able to use the framework. By immersing themselves in the intricacies of multi-agent team dynamics, professionals can acquire the requisite knowledge and proficiency to thrive in today's collaborative work settings. Embracing online classes tailored to address the subtleties of forming effective multi-agent teams represents a proactive stride towards honing essential skills and keeping abreast in the ever-dynamic professional sphere.

Automatically creating new content from old

You know what I hate? Repetition. Still, a difficult lesson that I’ve leaned, through both experience and discussions with experts in marketing, is that repetition is the key to getting one’s message across. I like to believe that a single, coherent and cogent conveyance of thought is all it takes to deliver, for example, a new security model or change in policy, or whatever. In reality, the propagator of the message –in this case, the security team– has to deliver the message, announce it a second time, send an email, post a document on the intranet, have more meetings with the same people, make sure to cover the topic on various 1:1 meetings, etc. etc. ad infinitum.

And that is why I have been working with Daniel Miessler’s fabric framework to devise a prompt that can take variations of the same messaging and create new, yet fresh, versions of the same message. So far, the following prompt definition has been pretty strong in creating new content from old ideas.

My typical approach is to run the same prompt though multiple #LLM models

cat examples.md | fabric --model gemma2:latest -sp create_similar

Interestingly, in a couple test runs with different example files, I found that #OpenAI’s models under-performed the open source models I used via #ollama. To be clear, this is uncommon. Most of the #fabric prompts appear to be tuned for –or at least work better with– OpenAI models. With this particular prompt, this was not the case. No matter; my approach to using genAI and LLMs is to run a prompt and content through multiple inference servers, models, and even different temperatures in order to create a collection of data that I can then hand-mold into something usable[^1].

Fabric Prompt

Here is the system.md contents from the fabric “create_similar” prompt

# INPUT

INPUT:

# IDENTITY and PURPOSE

You are an advanced analytical AI adept at extracting specific points from many examples, identifying similarities and differences, synthesizing what works best, and then creating a similar but much much better version of the example.  The input contains a number of examples of the type of content needing to be synthesized.  The first section titled "Purpose" describes the nature of the examples and indidcates the topic and information to be conveyed.  Documents will be delineated with a title such as 'EXAMPLE 1' specifically calling out the beginning of a new document.  The end of each example is delineated with a footer such as 'EXAMPLE 1 END'.  Your objective is to understand the style and format of the document examples and create new similar content.

# STEPS

1. Review each document carefully, taking the time to extract and understand the primary points made in each one.
2. Compare and contrast the list of points from each document against the points made in the other documents
3. Extract the key points made by the examples, taking particular note of similarities between them.
4. Output the results as specified in the output instructions.


# OUTPUT INSTRUCTIONS

Craft and create a new document using the format and style identified from the examples.  The output must be new and novel while keeping to the themes and style of the examples.  Do not use any language from the examples verbatim.

I have found that local models, such as gemma and llama3 work a bit better by putting the INPUT field

Purpose

Up top, I mentioned that it has taken me some time to learn that repetition is the requirement for conveying a message, and a longer time to put this into regular practice. This goes 3x for our individual contributor team members. Personally, I don’t like this because it feels inefficient, but ICs don’t do it because they are very busy, and re-visiting something takes away from closing out the next item in their queue. Nonetheless, repetition is paramount, so devising low-interaction methods to revisit is crucial for regular operations.

Footnotes

[^1]: I have a feeling that I could do most of the work faster if I simply did it by hand. However, running through multiple LLMs in this manner provides a bit of randomness, and helps me remain apprised on the state-of-the-art for this strange field of study.