Let's Build: Automate ANYTHING With AI Agent Teams (Step-by-Step)
ฝัง
- เผยแพร่เมื่อ 29 ก.ย. 2024
- Let's build an AI agent team using CrewAI for multiple real-world use cases. In part 1, we'll set up the framework for taking in a list of URLs, having our agents scrape the websites for content, and then summarizing that content. This is something I need for my newsletter. In part 2, we'll expand on that use case to make our code more robust and production-ready.
Enjoy :)
Code: github.com/mbe...
Cloud IDE, Fine-tuning, Inference, Data Management and more!
Try Lightning AI Studio FREE Today: bit.ly/3uBiWO9
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewber...
Need AI Consulting? ✅
forwardfuture.ai/
Rent a GPU (MassedCompute) 🚀
bit.ly/matthew...
USE CODE "MatthewBerman" for 50% discount
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Media/Sponsorship Inquiries 📈
bit.ly/44TC45V
Disclosures
I'm an investor in CrewAI - วิทยาศาสตร์และเทคโนโลยี
Do you like this format of video?
What other use-cases should I cover?
Yes we like it
Love this, Matthew :)
Have you ever looked into docker for this work? I use it constantly and it helps me deploy local models + agent code to end user laptops (and servers ofc, but most of my work is HIPPA and privacy focused) 😊
If you ever want a deep dive for your content, happy to show you the ropes!
The tldr is to simply use a python docker container and bake your agent in there, and then using docker compose you can combine ollama container with GPU access + agent container. Makes it almost magically no config for the end user
While, what you have been doing creates bu!z and excitement, I truly believe we need to certainly classify, basic, intermediate level learning stuff somehow.
In basics, one must only discuss fundamentals and intermediate to provide a demo.
And I understand it takes blood to create such hour long videos.
But you are so ahead in the curve, you are best fit for the job. Guide and be the light! 🙏🙏🙏🙏
Bro keep this going; recently found myself thinking
that from all the channels on the topic
I trust your message the Most.
Feels true & Authentic.
Regarding the format,
Absolutely longer video is nice
Separates you from other
News reporters.
In fact, I just watched the first
1:27 minutes, stopped.
Gotta do the due Dilligence
and follow through tomorrow
with better circumstances.
Really wanna have the together coding experience without spoilers.
Sincere gratitude to you for the great work, Matthew
Peace to AIl
Ivan
Love the format. I was building out research agents recently using langchain. It's an incredible set of tools. Tooling is definitelt where the fun is at. Gorrila will take human input and spit out API calls.
The question is where to invest time? At the pace things getting bombarded in AI space, sometimes i feel, wait and let it mature but at the same time i have FOMO, Who else feels the same?
Invest time in your biggest pain points, at least with every new development, it can be applied to your biggest use case.
I feel this. Definitely had a tool the boss had me build become obsolete over night. But! I still think it was valuable to do. Lots of techniques I learned along the way and getting comfortable with these "squishy" ways to get what you want is very valuable and transferrable.
Yes.
@aimademerich my friend in my 40s where I have achieved financial independence, actually there are no pain points, I just love to keep myself hands-on on latest technologies. But I have not seen anything like this.. There is hardly any structure visible atleast from learning perspective.
@@keithhanson9441 Seems like get your things done is the mantra :-).
Matt, again, thank you for all of this!
if you are going to extract a specific element from the html perhaps the websearch tool is what you should have used
How do you get that green circle on your mouse pointer? It looks useful for reading
This is sort of not really agentic more prompt chains I guess you could say.
Good for me. ^^
❤
Man I like your content, but you should really learn the fundamentals of Python instead of doing trial and error. It will save you a ton of time in the long run and and will make the whole experience much more enjoyable to you and to the viewers.
Nah bro, this is how we learn from his mistakes in a fraction of the time.
@@anubisai That will be a terribly inefficient and frustrating way to learn. You need strong fundamentals on which you can build upon
Love this format - You took us all step by step including "mistakes" YAY. This is how we all learn. The mistakes not edited out makes it real and approachable. Thank You for being so helpful!
Ah ya!! 👊😎
Hello Matthew. I'm Brazilian, I live in the interior of the state of Paraná, I'm amazed by your content. Here in Brazil, particularly me, there are few people talking and explaining about the creation of AI agents by CrewAI. In fact, only a select group of enthusiasts, followers of Alan Nicolas, have the knowledge to explain this topic. But they themselves are in search of knowledge. However, it is paid content. And not everyone has access to the basics. I would like to express my gratitude for your innovative initiative. I don't know anything about English, but I'm trying my best here to learn. I use AI and Google translate, it takes time, but I think it's worth it. Please continue to make this content available, your teaching is excellent. If possible, teach me the basics and fundamentals, with your videos. Gratitude
Amigão... sou do Paraná também, entusiasta da AI, se tiver algum grupo ou algo do genero, me informe, detalhe que acho interessante é que o desenvolvedor do CrewAI aparentemente é Brasileiro....
olha os BR, sou do RS
vamos criar um grupo ou algo do genero para trocar conhecimento... @@joneslemos
Olá Milton, sou de Curitiba, se quiser trocar idéias, bora se conectar!!!
Brasil!!! kkk Bora aprender!
Automating my life sounds awesome but I'll probably just end up more lazy.
I love this format. Could you put chapters in so it's easier to jump to the parts I'm actually using and skipping over stuff I already know? Thanks again! Great work.
Just want to share that Iove that you not only cover the news but you actually show us the practitioner side.
I share your videos with non-tech fam and they really appreciate your breakdowns, and you save me a lot of words 😂
This is awesome content!
Great video, but FYI: "RC" in a semantic version number stands for "Release Candidate" not "Release Client".
This is the best example of how development is done in the age of AI I've found on YT. I like that you said you are not entirely confident in your Python skills because I'm sure a lot of us can relate; AI brought in a lot of new coders (like myself), and we're all crash coursing our way to DIY AGI.
What's more important than the final code was your chain of thought through problem solving. While most of us won't be able to FaceTime the founders of tools we're using 😜 you went to ChatGPT for the scraper. I love this format, I love the thought process, I'm always stoked to see a new video.
Live editing is a great idea! I know you like to be all together, but it would be fun and useful to debug with you
You are phenomenal for this one!!! Thank you!! That IDE inclusion with this tut is everything
It’s literally GOLD. I’ve been thinking a whole day how to build it with langchain, exactly the same use case. I want to automate newsletters. Initially I created it with notion, zapier and chatGPT, but want first code approach. It’s the sign . Awesome stuff
Man, I applaud you for this. The vulnerability is real. And it makes me much less anxious of diving deeper into code. Especially with lightning, the hassle of an environment setup and some stuff that just not works on my local Mac, is not importatnt anymore. Great stuff!
It is well doing, seeing another one, playing around. This is motivating.
Good video we like it keep it up upload videos like this man lit 🔥
I absolute love this video. These types of videos will double your subscribers by the summer, if not more. I promise you.
I'll never understand the roleplaying in crewai prompts. Zero benefit to telling an ai it works at a top firm, it has no idea what that means to the context of the task. Best case scenario it just ignores it as noise, most likely it muddles your prompt and causes reduced performance of some kind.
"You are a researcher at a top firm, use your phd in data science to call the tavily search tool with {search_term}"😂
can anyone help me find the documentation?
thanks!
Thank you for this content! I’m really excited to dive into this
Thank you for the clarity. All the best.
❤I can’t thank you enough for this format as well as the high-value content of all of your past videos. 🎉 As you already know, this format benefits three parties: developers can learn new stuff as you deep deeper with your tech expertise, non- tech entrepreneurs can catch up with new tech trend with even some understanding, so you will not only attract existing audience but also people who wouldn’t take a lot of time on your channel otherwise. Sorry for the obvious! 😅
You are a very good teacher. That's the way coding works.
Looks great. It also looks like programmers are going to have jobs for a while longer. They are just going to need to retrain for new architechures, methods, and syntax. Of course new AIs will be there to help them transition.
Imagine how well the process of getting your CrewAI app to work would flow if AIs could reason.
Yes, ChatGPT wasn't even able to help Mat out. AI will need some further improvements to replace coders.
The agents will be in the thousands and finetuned for specific tasks, the Tool addons I will make will be A LOT. - I will make them as the human mind works and have an agent for each. reasoning, ideas development, and the prompts used will be Extreme, the best part about it is the AI will help make those personas and tools and its job will be to drop them in the folder and the program will "refresh" and the new tools added, and you can say with this list of 1k tools, I want to do this tasks. X what is the best way to do it with those tools, and then let the AI go to work and delegate, and assembly line certain content.
🎯 Key Takeaways for quick navigation:
00:00 *🏗️ Introduction to building AI teams for task accomplishment using Crew AI and Lightning AI in the cloud.*
00:27 *🔧 Explaining the use of various tools like Lang chain and custom tools in Crew AI.*
01:10 *💻 Demonstration of setting up a coding environment in the cloud using Lightning AI.*
02:16 *🤖 Initial setup for using ChatGPT with Crew AI and managing API keys.*
03:25 *🛠️ Correcting an installation mistake to ensure the latest version of Crew AI is used.*
05:12 *🗞️ Creating a Crew AI task for summarizing AI news articles for a newsletter.*
07:48 *🔄 Adjusting agent roles in Crew AI for more efficient task delegation.*
10:33 *🔍 Tweaking the scraper agent's task to more effectively retrieve and pass content.*
13:52 *🚀 Initiating a test run of the Crew AI setup and troubleshooting errors.*
16:25 *🛠️ Resolving import errors and fine-tuning the code for proper execution.*
20:30 *🤔 Adjusting the code to allow the Crew AI to receive and process multiple URLs.*
24:57 *✏️ Modifying the task code to effectively pass URLs to the agents for processing.*
26:48 *🔄 Realizing a mistake in the code related to the instantiation of the newsletter crew object.*
27:57 *🐍 Sharing personal experience in Python programming and encountering various errors while building the AI application.*
28:40 *🤖 Identifying a missing kickoff command in the newsletter crew's run process.*
29:26 *⚠️ Encountering a limitation where the AI tool cannot directly access and scrape website data.*
31:16 *🛠️ Consulting with the founder of Crew AI for troubleshooting and guidance on the project.*
32:51 *👨💻 Using string interpolation with Python variables to correctly pass URLs into AI tasks.*
34:05 *💡 Discussing potential modifications and additional use cases for the AI crew application.*
35:29 *🔨 Determining to create a custom tool for web scraping to fulfill project requirements.*
37:25 *🧰 Starting the process of developing a custom web scraping tool for the AI project.*
39:23 *🌐 Implementing web scraping in Python using libraries like requests and BeautifulSoup.*
43:10 *🛠️ Troubleshooting import errors and refining the custom scraping tool's functionality.*
46:00 *🖥️ Modifying the web scraping tool to mimic a browser for successful content retrieval.*
48:13 *📉 Encountering issues with websites blocking scraping requests and exploring alternative solutions.*
49:54 *🆗 Successfully scraping content from a different website but facing a new issue with message length limits.*
51:22 *🔎 Using HTML inspection to target specific content on web pages for scraping.*
52:31 *💻 Adjusting the scraping tool to work for a specific website, acknowledging the need for a more general solution in the future.*
53:26 *✅ Successfully scraping website content and obtaining a brief summary.*
53:57 *🛠️ Encountering an error with the scraper tool, indicating a limitation in its current implementation.*
54:24 *📃 Receiving an empty result from the AI agents, but recognizing the successful extraction and summarization of the article.*
54:37 *🚀 Considering enhancements to the AI agent system in future work and appreciating the capabilities of cloud-based development with Lightning AI.*
Made with HARPA AI
00:02 Setting up CrewAI for task automation with native and custom tools.
02:04 Setting up and using CrewAI for real-world tasks
07:04 Automating real-world tasks with CrewAI
10:18 Task 1: Scrape and provide content to writer
16:22 Importing tools from CrewAI and exploring examples.
18:56 Automating real-world tasks with Python code
24:33 Automating task by passing it into the task company.
27:11 Troubleshooting and resolving errors in Python code
31:51 Struggling to pass URLs into stock analysis tool
34:05 Discussing the use of website search tool and potential need for website scraping
38:34 Creating a website scraping tool using Python
41:01 Integrating code and addressing potential issues
46:04 Using third party tools and services to scrape websites
48:40 Struggling with complex scraping blockers and open AI's maximum context length issue
54:05 Lightning AI makes tasks easy in the cloud.
Analyzing step-by-step automation with AI agent units is truly charming! 🤖 Keen to dive into directing multiple AI agents seamlessly for various applications. This approach could change many industries. What are your thoughts on using multi-agent techniques in automation
This was fantastic! I have written a webscrapper with AutoGen and it was a total b*^ch... It made me feel better that I wasnt the only one who has had some difficulty getting one functioning.
Bro I have no clue how to code, so I'm clueless and lost watching these AI agent tutorials. Is zapier or bubble no code the best route for me?
Trying to follow you. Never used all this software before. Took me some time to figure out what you use. I now have Visual studio Code + python 3.12 and 3.11.8, but crewai0.14.0rc cannot be installed (wrong python version). installed most recent crewai, but on execution of the script it cannot find modue crewai. Path seems correct and pip list gives it. Stuck here.
This format is great. Sure you wouldn't want every video to be a long form tutorial but this is really helpful and is great when interspersed with your informative and, ahem, "shocking" ai news content. 😉
I'm enjoying the video. But during the middle, I want to type and ask you a question about why you did a specific thing. So, what if you streamed this? You could go through something like this with viewers asking questions along the way. That would be great. Also, could create more views for the channel -- it would also be advertising for your AI consulting. Could be cool.
You are the only one of about 5 "crewai for dummies" videos that has even mentioned the IDE used. In this case its lightning AI, only trouble is its taking days to verify my emailsince I dont have an edu or org address, who knows, it may even not be verified. What other suitable cloud/local IDEs are recommended for developing crewAI projects?
Great video Matthew! When trying to hook up langchain/groq instead of OpenAI I ran into an issue. Even though I do not call it at all, I am asked for an OpenAI API key. I only have a free account at OpenAI at the moment, plus I don't even want to use OpenAI. Did you run into this issue as well when hooking it up to a different LLM?
Hi Matthew, thanks for posting this video, very helpful. I'm interested in using agents and want to invest some time in a framework but don't understand enough (yet) about the different frameworks. Autogen is the other framework that gets a lot of mentions. Could you please help me understand why I would choose CrewAI over Autogen and visa versa? Thanks in advance.
you make a spelling mistake in the title: "Shocking CrewAI Tutorial - Automate REAL WORLD Tasks From Scratch!"
thinktahnker as scraper kinda gives a false sence of scraper i my mind and more information masher. cant describe it but yea there was i saw first, its like yea he invent fake urls. im gonna try this my self and mix, y for video. and might be worth mentioning if im not mistaken not all sites are giving you allowence to scrape and, you can scrape but i would give him a simpler description. i use anotherway to get more correct data, again not sure if yea, but i scaped manualy and learned meta data also you can use 2 methos now days or token problem but also cant we just make more agents for that like caching a big agents into smaller like the models ? or as another video i watched. the implemented agents who dont neede to remember deletes its pre-que-dir-after-sourced/resets something And have to add yes on the lastest request on live, but time zone diff for me.
It seems my error is with the Scraper Tool as I get "Article with specified ID not found." Looking at the code of that tool used I see line 20 (article = soup.find(id='insertArticle") is the likely source of the issue
This is totally learning by doing and it's the absolute best way to learn! The way you show us your debugging process and how to solve every problem you encounter that is so incredibly valueble.. Please keep making this type of content
Matthew, thanks a lot for this extensive demo. Seeing what can go wrong and seeing how you solve the problems was absolutely inspiring. Thanks!!!
I have but a single ask…
I know they’re essentially edge-docker containers.
But please use a venv or pdm or something, at minimum to prevent dummies from globally installing something somewhere else and completely stuffing their os.
And well… the use case here is pdm generates a file struct for you and makes it wayyy easier to work with
I want to see live editing.
Thanks Matthew for the great content. Please, do you have some step by step guide as I have some difficulties in finding the correct text to copy into the documentation? Kind regards,
You are great man, you gave me confidence, I've always taught that I lack to much in my python skills but seeing you struggling with it to and still making it work gave me confidence!
Hey Matthew, this is THE format. Like my maths teacher always said: you always get more credit when you show your “working out”…because it demonstrates how you reached your answer.
wow, this tutorial was very bad. Novice at python that just decides halfway to copy another example without knowing how it really works - isn't that what we're all trying to avoid by us looking for a tutorial?
You have a great style of video. The problem for me at least is langchain, crewai, and a number of other dependencies are changing so fast, everything is kind of useless to me. Minor version of a lot of libraries are completely incompatible and langchain especially is constantly restructuring their own libraries so imports break.
Loved it, no complaints, we live in mistakes, its good to see them, it helps eliminate some of the "fear of starting". ...However, I wonder if there is some way you can do a video that is good to listen to on dog walks .. Something that does not require us watching the screen (e.g. Interview). Maybe more theory .. That would be a good "mix up", and are fine to be long.
YES! Love everything about this especially using Lightning AI and developing a solution from end to end that actually solves real problems.
I must have missed something. How does CrewAI differ from AutoGen , AutoGPT, etc.?
Yes, I like this. Going through exactly the same kind of thing we are all doing right now.
I just give a thumbs up automatically on new content before clicking play. I know it’s gonna be good anyways
🤫 Good one. Matt, thank you for showing how AI works. Keep going. 🌹👏👏
There is no better form than showing all the trials and mistakes of trying to make something work. Because if you edit out all the mistakes... no one really learns anything unless it goes perfectly. Kudos.
Funny part is I got stuck in the same part but I misplaced the Founders #. Thanks for helping with that 😂!
The scraper thing is a problem I ran into. There are a lot of scrapers out there, but they all seem oriented toward populating massive datasets and not for summarization or other uses. I ended up having to write a scraper in python. It was a pain.
i really enjoyed this!!! screaming at the screen self. self.. self.... then dont use the same variable names!!! don't dony dont ..... ahaha namespace pollution
You are doing great job you update us with AI I hope you get 1 million subscribers very soon 🌹
I like your style buddy.. you explain things pretty good and you have a large variety on your channel. You look a little bit like pauly shore
VS Code can also be set to save automatically, so there will be no need for CTRL+S :)
Could this be used as an email scraper? Like being able to define parameters for scouring through years of accumulated email to get rid of the crap and keep the stuff that isn't crap?
Did you come up with a more generic web scraper tool. The version in your git hub is still hardcoded to work with that one site.
I wanted to create a bot that does this and scrapes real estate news in Bangalore and makes posts and phblishes to social media
The coding format, with all the re-works is awesome!
Ironic you didn't use something like co-pilot to help you resolve the code errors..
Fake video, what I mean is he is trying to act like he is coding this on the spot but he is not, good content should just be honest about it
Can crew ai control other models like autogpt or self operating computer?
Do you like crewAI over AutoGen? Which would you recommend ?
Which one is your fav Matthew. Autogen or Crew AI and why?
Just been reading crewai docs, a PDF write tool would be a good addition to the toolset.
Thanks so much, this was exactly what I was looking for!
Hello sir can you make a tutorial on how to build the crew ai agents with llama2 or llama3 local llms
I really appreciate this video. Thank you Sir!
Bro called the founder to show him how to do string interpolation xD
Real interesting ai use cases are in neurosymbolic.
just imagine leaving all these debugging to AI….
does it work with any other API other than OpenAi's
No hate, but the only use case anyone has dreamt up for AI agents so far is ... "Take a bunch of things and summarise them". What else are we going to use it for? 🤔
Haha!
I guess u missed the part of the video where AI generated all the code ;)
Please review open code interpreter. Thank you!
📝 Summary of Key Points:
📌 The video demonstrates the process of building a crew AI to accomplish tasks together, focusing on tools to extract maximum value from allowing agents to use tools.
🧐 The presenter sets up a team of agents, including a scraper agent and a summarizer agent, to scrape website content and provide summaries, showcasing the use of custom tools and collaboration in a cloud IDE.
💡 Additional Insights and Observations:
💬 Quotable Moments: The presenter showcases the use of Lightning AI, a cloud IDE, to code and collaborate on AI applications, highlighting the benefits of real-time assistance and cloud-based development.
📊 Data and Statistics: The video demonstrates the installation of tools, creation of custom tools like a website scraper, and the delegation of tasks within the crew AI setup.
📣 Concluding Remarks:
The video provides a detailed walkthrough of setting up a crew AI, utilizing tools, and customizing agents to perform specific tasks like scraping website content and summarizing articles. The presenter's use of Lightning AI showcases the power of cloud-based development and collaboration in building AI applications.
Generated using TalkBud
Thank you for sharing your thoughts process.
Have the same use case (weekly AI newsletter in German): currently I'm using a Perplexity collection with custom instruction set to create a summary of a website. Perplexity does a great job on this as I'm not dealing with token issues or blocking scrappers. But the Perplexity API doesn't support calling my collection "GPT". So the process here is broken.
Looking forward to see your progress with this use case 😊
I had SO many issues with web scraping, I was practically shouting at the screen "No, not that way!" 😆 It's one of those use cases where with experience you know what will and won't work and exactly how brittle your solution will end up whatever approach you finally rest on - I think it's one of the best arguments for an R1-type model rather than an LLM getting fed source code etc. It raises the point... will all these web platforms ignore our intentional breach of their TOS via automation? Seems like a 'last stand' nobody's talking about yet... The scarping and 'brittleness' issue is one reason why I'm currently leaning toward using Autogen Studio and their 'skills' approach instead - it might be okay for the skills code to be brittle and get evolved when it inevitably breaks, with that kind of modularity built in, much like how e.g. Zapier updates the tehcnicalities of how their building blocks work in the background as issues arise, and your automations/flows don't need changing fundamentally (usually!)
you can do this with pretty much any coding language
Agent is just a hype buzz words .
I have been working on similar AI code and getting mixed results. The struggle is real, keep up the good work.
why don't crew write code for itself from text?
You said you'd share the code. Where can I find it?
Thanks for all your work!
Oh shoot! I forgot. I’ll share it now.
Huge Thank You!
first of all I'm an avid follower of your videos and your Presentation style is superb. Regarding agents this just seems like a very convoluted method of achieving the same thing with gpt4 in a well worded prompt, I don't need an agent?
Great and thanks for this demo that shows also the pitfalls and how to get rid of them. Did you know that VSCode has the "Live Share" plugin (Microsoft) that allows co-editing in our local vscode, with others in using their vscode? Same as what you shown with this cloud based editor (good to know this one too).
Can you so a video showing how to do this in flowise?
Great video !!!! really helpfull
They show the first year old is that the website blocks IP originating from a well known hosting/cloud. Running from your computer should work.
"Bill.... i've got a problem getting my windows to update..."
Great video.... this format is SO MUCH APPRECIATED as the errors you get highlight the things we all come across trying to actually create our own projects. Your troubleshooting is what we really need as examples.... not 'now pip install, blah blah' that most creators videos are.
Thank you for trying this style out and please consider doing more of them, even if its just your version of trying to duplicate some new tech Repo's example project.
Great vid!
As a brand new python n00b, it's inspiring to see somebody more experienced than me troubleshooting through new code they're writing.
I really appreciate everything you're showing us. Thank you!