The Complete Guide to Building AI Agents for Beginners 

Got it!

+1

+2

Found this channel very late I guess. As a commerce student I'm eager to hop into this journey. Can you pls breakdown on how to get to this role with the specific tools, packages and frameworks to learn ? Thanks a ton

It also means that technical skills, while essential, will need to be paired with a vast life experience if one wants to become a top performer as an AI Agent Developer.

What's Ur email let's talk business

It's great that your company is developing it, how did you do this using GPT-4?

Yep. The insights and value are knowing what questions to ask. Tech does replace humanity, it supplements and compliments it...nice work by you

@@PedroHenrique-cu1ur with a lot of c&p, try and error, frustration and passion, I would say. 😂 The road to the goal is paved with failures. There were a few moments where I thought I wouldn't be able to do it because I couldn't solve a problem, but so far I've always managed to make a breakthrough, even if it sometimes took weeks.

@@user-vb3eg7mt2li’m doing an ai agent project as we speak but i have zero experience with coding and development 😂😂 but im pushing through. I need my own assistant.😂😂

Yup, working on it

which one did you like more between crewai and autogen ?

@@amrinderrandhawa6799i clearly prefer crewai for the very customizable workflows and the ability to not only assign one, but two separate llm's to a single agent if necessary. for me, autogen was more of a "quick and dirty" solution to simpler tasks, but if you need a series of tasks relying on each others outcome and each task requires about 4 agents to be solved properly, you quickly end up with 20+ agents divides into 5+ crews and 5+ tasks to solve, where the quite straightforward "question - response" style of autogen just does not really do it anymore and usually gets pretty messy, as well.

@@amrinderrandhawa6799 clearly crewai; i tried autogen first (mostly because at the time i tried autogen, crewai has not been released yet), but working with autogen feels a lot like "chaos management", whereas crewai is way more focussed, more streamlined and more intuitive imo... both work pretty smoothly if you get your setup going (for autogen, i used olllama and litellm to host multiple llms in parallel on my local devices), which also works with crewai even without the intermediate step of using litellm and both offer very similar capabilities; some of the differences are: autogen offers way more flexibility for agent creation with the "agent class" parameter, which allows to add (to put it videogame-terminology) some sort of "passive skills" - if the ability would be the "active skills" in that regard. - talking about agent classes, one of the most useful classes (at lest for me) would be the "open ai assistant api agent"-class, which can be used as a drop-in replacement for the openai assistant api - which might come in handy depending on your use-case -crewai also offers something that might also come in handy, depending on your use-case: where autogen "only" offers one parameter per agent to set a base url and thus, assign your desired llm to it, crewai actually offers two, the "llm" parameter and the "function calling llm" parameter, enabling each agent to be assigned to two different base urls, if necessary. crewai offers you to define crews and tasks, not only agents, which for me, is the dealbreaker here. as i only (!) use sequential processing without asynchronous tasks, the ability to not only create the agents, but then assign them to different crews, those different crews to tasks and those tasks to processes makes it very "applicable" to real-world use-cases by being very modular by nature. maybe i was just unlucky or i did it wrong, but when you try to make multiple agent-teams interact with each other with them relying on each others outcome, autogen never provided really decent results.... -as stated in the video, as crewai has multiple natural language prompts in the backend, the outcomes might vary depending on your prompt engineering capabilties and (mostly if you want to use crewai with open source llms and not openai) the natural language processing capabilities of the underlying model, as well. (sorry for the wall of the huge wall of text... 😅)

@@amrinderrandhawa6799i prefer crewai over autogen mostly because it focuses more on crews and tasks, whereas autogen seems more agent-focussed (at least, right now...). maybe i just got unlucky or maybe i was doing it wrong, but bigger the agent swarm grew and the more complex the tasks got, the worse autogen performed.

hey i would like to know ,can i create this crew agent & all completely free ??most of them need some sort of Open ai API or some paid services?

Agreed

What I really like about agency swarm from a conceptual standpoint is We can elevate an organizational model of how groups interact and use the framework to model how scientist and cognitive researchers think the human mind works Imagine if we can create agents that are essentially a collection of various processes that mimic what we think happens within people. We have conscious awareness, subconscious processes, and autonomous systems. All working underneath. The consciousness of our awareness yet are happening regardless. This sort of structure and thinking of the system can easily be extrapolated and presented as an entirely different use case outside what Arsen originally planned his solution to be.

Hard to customize because there are too many unnecessary abstractions for features that can be implemented by yourself in 5 minutes

ChatGPT

@@vrsen I solved some problems using it, but I could not find how to specify the model to use.

'readline' => i also had this problem, I fixed it with: pip install pyreadline3 ...on/for windows. And: import collections collections.Callable = collections.abc.Callable ...in python code

👍

Hey, I am also located in germany and would love to connect regarding this problem, how may I reach out to you?

If you deploy on Azure, all you data stays on your own private instance. vrsen.github.io/agency-swarm/advanced-usage/azure-openai/

@@vrsen thanks I will check that out. In the gov space this is a huge dealbreaker. If using Azure OpenAI, I also see why it doesn't make sense to support open-source models. Thanks

​@@playboy71322why his framework doesn't support open source models and what makes it different from crewai?

Watch the video.

You got it!

This is actually going to be the case after OpenAI releases memory feature in Assistants API.

@@vrsen its gonna get crazy 😎

pip install worked but agency-swarm genesis gave some issues, according to gpt it seems like there is still an issue with the readline module, specifically related to compatibility issues with Python 3.10. The error message indicates that there is an attribute error in the py3k_compat.py file of the pyreadline package. This issue might be due to the pyreadline package not being fully compatible with Python 3.10 yet. Which version of python is compatible? i've so far tried 3.9, 3.10.10 3.11 and latest with no luck. Would be cool to be able to get past the second step😂

If you want to use gpt-4 you need to buy openai credits

Watch the video.

Local LLM is hard though. No direct documentation, but you can learn how to use model in local.

the developer clearly states, no open source models, only OpenAI ChatGPT models. OpenAi does not make any production open source free models.

@@drnotebook he came out with a video shortly after this one saying that he changed his mind and in fact made it so you CAN use offline LLMs

Yes you can use fine tuned models on your openai account by changing model parameter in your Agent classes

@@vrsen Thank you!

visual studio code

can you elaborate on why? not challenging, just a noob wanting to hear another perspective

​@@outsidesoundsOpenAI is not open 😭🤣, why you ask that? Everyone know if you want to use OpenAI api you need to spend money

the point the guy in the video makes is that open source is not free, there’s no api fees but you still have to pay for your own hardware/infrastructure to run something at scale. is that not true?

@@outsidesounds open source mean the source code or source data is free. Meanwhile OpenAI never reveal which data they are using.

Free Open Source Software (FOSS) IS FREE. OpenAI is not FOSS

You can import tools form langchain, see “advanced usage” from documentation. Although, obviously, I do not recommend this. Use tool creator gpt instead

@@vrsen advanced usage from agencyswarm or langchain ?. Btw can you list your reasons for agency swarm not being llm agnostic in a seperate video. going into the nitty gritty details. i k you covered some reasons for that , but I think people would like to know more about this. And maybe some ways to make this work with other llms

I think this video and the previous ones sufficiently explain why the bet was made on the Assistants API, particularly in terms of state management. And within 3-4 months, OpenAI will release a new model, making the agency swarm both cheaper and better🎉

I totally disagree. In my opinion OpenAI is at least 6-18 months ahead of everyone.

the developer clearly states, no open source models, only OpenAI ChatGPT models. OpenAi does not make any production open source free models.

If by trained you mean actual training, like adjusting the weights, than It’s not trained. They are using publicly available llms like Claude and GPT 4. I will try to beat them in a few videos on SWE.

@@vrsen Nope. They are not using out-of-the-box LLMs.

the request model didn't exist

What do u mean?😂

pip install agency-swarm 🤣

Bro have you tested devin yourself? Only a handful of people has access to it it… And the videos they selected weren’t just some random performances, I’ll tell you that. My goal with this video was to help you get started. To make these agencies work well in production does require work. This is basically what businesses are paying us for. If you want everything to be done for you by AI, you’re probably around a decade early.

you need to investigate fine-tuning. it does require work and investment.

@@vrsen your system is world class, no doubt. Way better then crew AI and I love your videos, I was just saying that in practice it's that GPT4 is not yet up to the right level for these use cases. However, as correctly mentioned of the models are fine tuned also then it could work.

I have built a few apps including one for crypto and I have no coding experience at all, whenever I have an issue I simply use ChatGPT to help me fix it, that is what it designed for. This however is a little more complex and this is not for beginners as the title suggests, or maybe some type of beginner that I dont know of :)

out of my personal experience, also heavily lacking coding-skills: i ask the ai to fix it for me ;)

I have built many apps with AI and have zero coding experience I simply use ChatGPT to fix the issues when they arise, that is what she is designed for. However this is way more complex to fix and to be honest the title suggests it is for beginners, but beginners of what? it is more of an info video not a complete guide. For a beginner you would need to slow down 500%.

AI can already help you fix the majority of the issues; the important part is knowing the right questions to ask. I will release a video on AI Agent Developer skills soon I guess, since there are a lot of requests.

The AI finds and does the fixes. I know nothing but have been quite successful running python scripts.

what is that