Master CrewAI: Your Ultimate Beginner's Guide!
CrewAI Colab: drp.li/OeAmK - CODE
In this video, we'll be exploring an AI framework called Crew AI that empowers agents to collaborate and work together seamlessly through role-playing. Think of it as orchestrating an efficient and cohesive crew of AI agents, each with their own unique roles and goals.
🕵️ Interested in building LLM Agents? Fill out the form below
Building LLM Agents Form: drp.li/dIMes
👨💻Github:
github.com/samwit/langchain-t... (updated)
git hub.com/samwit/llm-tutorials
⏱️Time Stamps:
00:00 Video Overview
00:30 CrewAI Intro
01:13 What is a Good Agent?
01:53 What do you need to create a good agents?
02:37 3 Components of a Good Agent
02:44 A Good Large Language Models
05:02 Good Tools
08:12 A good agent framework
14:29 General Overview of CrewAI
17:27 CrewAI Core Components: Agents
24:47 CrewAI Core Components: Tasks
30:10 CrewAI Core Components: Tools
34:10 CrewAI Core Components: Processes
Sam you did it again with this one. Single handed leap forward in understanding. Thank you sir.
Glad it was helpful!
Thanks for taking the time to make a polished video explaining agents in detail!
Thanks for making this! In terms of the intro, for each clip I think making it a bullet point on a slide would be the way to go. Stay positive and keep up the good work! Subscribed!
Just a note in case anyone hears it and is confused/concerned: when configuring your agents, if you don't define an LLM, it won't fall back to using GPT4, it will fall back to using the default model that's defined. Thus, if you define your default model as gemma-2b-it running locally through LM Studio, the agent will fall back to that, not GPT4. If you don't define a base model though, then yes all you agents will default to GPT4. Thanks for the video, filled in some spots that others missed.
Hi thanks for pointing that out. I just burn through 5$ and the agent hang up after running through a loop. Obviously it wasn't converging towards the goals. Eventhough Crewai recommend GPT4, I would like to try it with mistral running at Groq. any pointers for that will be appreciated.
Amazing tutorial! Starting my journey in agents and your crew AI tutorial has been the holy grail for me!!
excellent overview, thank you. what a time to be alive!
Your credentials are impressive Sam. Everyone benefits from you and your teams wellspring of knowledge!
Thanks for the kind words.
@Sam, thank you very much for a very well-structured explanation of CrewAI and the Agents.
Just got started with crewai yesterday and I must say this video is amazing for kickstarting newcomers like me
this is the second video I have seen on AI , Crew AI and helped me a lot to understand the concept
Thank you, Sam. This is very informative, and I appreciate the detailed and clear explanations and walk-through.
Thanks Sam. Waiting for the next ones.
The video overview was terrible. I thought the feed was broken
Agreed, the clips in the overview need to be longer so we have context, that said the video does look incredibly valuable so thank you Sam!
This is good feedback from both of you. thanks. First time we tried to do something like this. My fault for telling the editor to keep the overview very short. Thanks guys.
I’d say it could have been better, not horrible. That preview style fits better a podcast or interview. A better intro for a course imo is outlining the contents that will be covered.
It felt like the intro video was made by AI 😅
maybe like an intro video you speaking showing the overview then we go into the video would feel like a better flow for long videos like this@@samwitteveenai the issue was that you left the original audio and just used a fade on the audio making us the viewer think that something was wrong in our end... like a simple voice over that part of the video would easily fix this issue...
It’s worth noting that you can get surprisingly good results in specialized agents using modest models. You would have to use a LoRA, and Context Free Grammar to constraint its output to do proper function calling, but it can be done if you’re getting an agent to do a specific task.
Thanks for breaking down this in laymen terms. Very helpful video for AI newcomers!
looking forward too your next video on CrewAI
Very cool, navigating through these frameworks and options feels like I'm a kid in a candy store-overwhelmed but super excited! 🍭
I love your videos. Great balance of detail, theory, and practical examples. Looking forward to the more customized example video!
Thanks!!
Thanks Sam! Excellent work!
Magnificent, excellent post, for a while I wanted to see something as great as what you just explained in this video, thank you very much, greetings from Colombia
Great video! Thanks for putting this together - it's an excellent guide to building with CrewAI and the current state of agentic systems in general. Bookmarked for future reference.
Glad it was helpful!
Brilliant Vid! Just beginning to look at the AI dev tools, and this is really useful. Thank you!
fantastic intro! thank you!
Wonderful work, Sam!
I found the video to be very informative and I gained a lot of valuable knowledge from it. Thank you Sam.
Glad it was helpful!
Such a good and informative video, thanks a lot!
Super helpful and informative. Thank you.
Great video. I’m a big fan of crew Ai and hope this will encourage more people to try it out. Combining crewai with fixie ai agents looks like a really interesting concept for building complex systems
Crewai is such a bloat of dependency breaking hell. How do you navigate it without creating isolated crewai only projects? How does it compare to Autogen and Taskweaver? You might see Autogen as more of a scaffolding framework but Taskweaver has more integration out of the gate. Have you used Taskweaver? I’m genuinely curious because whenever I sit down to give crewai a chance, I get blasted with dependency issues. Right now it won’t install with a clean project and LangChain because crewai hardcorded python-dotenv 1.0.0. You fix that and you’ll run into issues with unstructured, both being in their examples. Then you go to the repo issues and you’ll find mods barking at people confused with their dependencies. I just want to know from someone who really likes it, if it’s worth the headache over the other 2 that already work great. Is there some feature that just makes it amazing, some delegation functionality that’s over the top?
Good work on this Sam
Thanks!
Beautiful well presented concise walk through. I'm about half way through (on gf's acct) and look fwd to jumping back in later. Super happy to have found you!!
Glad it was helpful!
Just what I needed to learn about CrewAI without having to read all the docs. Great for visual people such as myself haha
Hands down, Easily the best LLM creator on Planet Earth
Intro had me “Hallucinating”
very nice introduction to crewai and the associated jupyter notebook. all works without any mess, i got when i tried to run it through vs code as was suggested by other tutorials. missing packages, conflicting python versions etc made my life miserable. I know now how to go around all that messiness but your code in the form of jupyter notebook is the best learning resource I think. and I am a retired prof who made living by teaching engineering. Just know enough about coding to get around different codes. since we started with fortran coding. never did any coding in python or exposed to object oriented programming that came afterwards.
Thank you Sam - Great video. Looking forward to more of the same getting into deeper topics like creating own tools etc. Something I'd be interested in, is creating a tool to connect to a DB and perform queries while understanding security. i.e. how safe is 'own private data' as it moves through these tools etc. Thanks for all your efforts and making these videos
Yeah I agree totaly about the safety bit. most agents that do use any kind of DB don't lock it down well.
I'd recommend having a look at the FOSS framework Tasking AI. A good start with a similar feature set that you find in OpenAI Assistant APIs, but with a "bring your own model" strategy.
How do you actually see the raw request and response from the LLM to debug and improve the prompts? I can see the model response but if I could understand what the compiled request looked like at each step I feel I could use the 18 months of experience in creating prompts to optimise. Particularly important for the smaller models run locally via Ollama
Love the video footage, what stock video repository did you use?
We are mostly just using what is in descript last time I checked.
One question I always have with agents is how granular you should have them be? If they are too specific and granular, as advocated in the video, then what's the difference between an agent vs a tool or chained tools?
The best to ever do it ayo?
Quite lengthy video, but it is a good overview of CrewAI
Great video ! However, I would like to know if you have expolored "AutoGen" and how do you compared it with "CrewAI"?
Yes I will make some AutoGen vids soon and will talk more then. I have written some comments comparing them in the CrewAI vids as well.
@@samwitteveenai Thanks will go through the comments. Also I feel that AutoGen makes more sense than CrewAI because of its Human in loop feature.
If my app is heavy on the prompting but not too much on the agents it's an ask and answer agent should i go with crewai or langchain(langsmith...)?
You could try CrewAI or AutoGen, which is more suited to just straight conversations etc.
Curious about the difference compared to metagpt?
There seems to be an error in the code for the tools, specifically get_contents. The tool should pass ids as list[str]. Without this change, the tool doesn't seem to be extract any contents, result in an error. It might also be a good idea to remove the eval() statement in the function.
CrewAI or Autogen - pros and cons? Should I just learn both and decide for myself?
I have videos coming for AutoGen as well will talk more about differences there. CrewAI is easier for most people to get started
did you tried the crewAi with claude 3 opus api? i think it could be really good. lot better then gpt-4
Yes but I have mostly been using LangGraph with Claude models.
Is it possible to build an app using CrewAi on Streamlit? If so, can you do a video to show us how?
yeah you can use Streamlit for the input and for showing outputs etc.
How much harder is learning lang chain than jumping into these solutions? I assume Lang chain is much more tunable and flexible, but what kind of learning investment is it going to take?
I don't think it is a huge task to learn, if you can't code etc then maybe yes but otherwise it should be ok
Can we use multiple crews in a single application? Let’s say one for finance and another totally different crew for travel planning?
Yeah you can set the up totally separate etc and have some other code call them.
It seems to me that crewai 's crew is only specialized to do one single task. and that task can be achieved by executing multiple tools. But we cannot assemble a team of crew(s) to achieve a multitude of problems. I think langchain agents can only achieve that @@samwitteveenai
CrewAI is just built in LangChain. You can have different parts of your app fire different crews, but I would say be careful using it for production
I’d like you to describe an “appMaker” that would incorporate already existing code in, say, JavaScript, rewrite it in Swift, design a UI as directed, and launch it on the platform I’m using. Maybe such already exists. Help, anyone?
What would be interesting is to see if there is any sort of proof that what comes out of this utterly convoluted way of programming applications is any better than if you did it normally. I can understand the argument that certain tasks like speech recognition, natural language processing or image generation cannot be done any other way, but what advantage does a bunch of half-witted "agents" have exactly against hardcoded control flow logic (some of which you apparently need anyway)?
Hello, my name is René and I need to implement a project for my agency. I am trying to find out if it is possible to create multiple crews in CrewAI and have them work together. My plan is to have a Manger LLM who then assigns the task to the appropriate crew. So I would have planned to create an IT department (crew), a legal department (crew), an analyst department (crew), and depending on the user's request, the manager assigns them to the right department (crew) and they collaborate with other crews. One of the tasks would be to analyse a PST file and search it according to certain criteria and create graphs of email traffic. What are your thoughts on this? Is it even possible to implement this?
Yes you could do this all in one crew but have sub managers who can delegate. Honestly probably better to constrain it more by using something like LangGraph
@@samwitteveenai Thank you very much for your quick reply. Sorry if I'm being cheeky, but do you have any tips on the best way to start? I know that you can have one manager for the agents, but multiple managers is new to me. Thank you and best regards from Austria
think of the main manager as like the CEO and other managers as like middle management below the CEO they are in between but they also can operate as department heads
Or maybe you should just use LLM as the "tool" or one of the many APIs employed within your non-LLM framework. Specifically, the tool for classifying and translating natural language into whatever domain-specific knowledge representation you need in order to let a traditional, hardcoded algorithm do whatever it needs to do. Where you are not limited by LLM foolishness and can focus on the task at hand, as has always been the case before LLMs.
After playing around with CrewAI for hours, I don't think it's the solution I'm looking for. As easy as it is to develop your own agents, and as many bugs as seem to exist in CrewAI, it seems like a royal waste of time. In the first version of CrewAI that I tried (the latest) the agents can't talk to each other. Rolling back to other versions yielded various other bugs in the framework that broke other things - the agents can communicate but can't use tools, random exceptions deep in the stack, and etcetera. It just doesn't seem very stable at all.
Verry didactic. Who said programming is dead?
Is crew ai free?
yeah it is an open source framework
that intro was super confusing and I thought there was something wrong with my video
I just tried crewai, partially because of this video and it's low quality has me flabbergasted. The documentation contains a lot of AI generated marketing fluff. The prompts that are sent to the LLMs are terribly written, full of grammatical errors and strange phrasings. They also make certain demands that steer the LLM away from certain types of responses like short summaries. The system does not support any types of prompt templates so using models that require specific prompt formats to work well is out of the question. Notably, that includes the new Llama 3 instruct models. Overall, terrible library that should be marked as a non-production ready alpha.
I totally agree about not putting it into production
Why does it take 17 minutes to get to the point of what an agent is...
This voice must be generated
Unfortunately The editing in this video is not the best I have to say. Sometimes it would just be easier to show a talking head rather than trying to illustrate something with stock / ai or repeating scrolling through sections of the docs videos that are distracting or cheap looking.
There is no "decision making" and "reasoning" going on in the LLMs whatsoever. There is just translation, imitation, copying, applying premade templates. Decision making and reasoning involves not only correctly constructing future scenarios, but also scoring them with respect to expected rewards, and taking into account actions of other actors. There is no publicly available LLM out there which performs that sort of work. There is a theoretical paper of how it should work (How to Do Things with Words, a Bayesian Approach), but it's a far cry from a working implementation.
lol in this one you are sounding like Gary Marcus. I get the point about reasoning and I don't think they reason in the same way humans do, but they are making decisions in way that would often be hard to code purely with heuristics. Thanks for the paper reference I will check it out.
@@samwitteveenai My point is: if I code a program which looks up a text template in a database which describes reasoning steps, and you run this program and the template is presented to you, you would not claim that the program was "reasoning". The fact that an LLM can at the same time present to you one claim and also an opposite (contradictory) claim, is proof enough that it is not "reasoning".
@@clray123 The LLM is not "reasoning," nor is it meant to. It's just picking what words come next given the current context. If the current context can lead to contradictory claims, then it is capable of giving either, or both. It doesn't have a "preference" except whatever "prejudice" has accidentally been introduced in the training data or current context. This is pretty much the same as with humans - limited context can lead to contradictory statements, as can prejudices in the "training data." For example, just think of all the religions/denominations based "solely" on the Bible, even the same version and translation, and all the variations between them, some of which are certainly contradictory. What frameworks like CrewAI are actually doing is just iteratively modifying the context (with external tools and different "personalities") until it can output a result that meets "what comes next" according to your criteria. Reasoning, also in terms of human thinking, is more of a "post-processing" effect - we start to say something (or go ahead and say it), then realize that it's incorrect, and correct the statement before giving it (again). Sometimes this process takes many iterations in humans as well. Here's a fun example to try, to understand a bit better maybe what I'm getting at here. I use this prompt: "When I say one, I mean two. I have one apple. How many apples do I have?" Now for this, llama3-8b will correctly respond with some variation of text that ends with me having two apples. Following this, I prompt: "If I get one more apple in the future, how many apples will I have?" And llama3-8b will now incorrectly respond that I will have 3 apples. The "weight" of my first sentence is now too low in the context stack, and it has "conflicting" possible options for "what comes next." After this incorrect response, however, I prompt again: "I said when I say one, I mean two." And now, it can correctly predict that I will have four apples, because that "important" context is refreshed. It's not "reasoning" anything, per se, except what words or phrases come next. It doesn't "know" what one or two even are, although it can provide the exact definitions for what numbers are and where "one" and "two" fall in numeric sequences. So, I would argue that although it isn't "reasoning" at all, its ability to present the emergent property of reason probably already far exceeds what you would find in the average human. My 6-year-old daughter can add and carry, subtract without carry, and do either of these operations on up to 3-digit numbers. She is easily capable of doing the necessary math, so I gave her the same prompts. She got the first prompt wrong. I had to re-prompt her twice about the "When I say one, I mean two" before she got it right. Then she still got the second prompt for adding "one" apple wrong, and had to be prompted again. llama3-8b was "more" capable of "reason" on this incredibly narrow scope than my daughter. She's certainly capable of "reasoning," just watch her planning how she can get away with something she's not supposed to do... she's clearly able to predict outcomes and adjust her behavior or words accordingly. But so can an advanced LLM, when provided with the appropriate context. With the iterative approach, this emergent "reasoning" becomes even stronger as the scope is honed down and the important bits of context are given more appropriate weights. Using "agents" with these "goals" and "backstories" sort of "freezes" part of the context on top of the stack, which gives it a bias in the "next words." Essentially, since it's able to give a "reasonable" answer, given sufficient context, even using a specific cognitive process (inductive vs. deductive reasoning, for example), I'm not entirely certain we can say that it's truly "incapable" of reason. It's (mostly) incapable of acting, but you're likewise technically incapable of acting in the one way it is able to act - you cannot directly add bits to electronic memory to be adapted for display, and you require use of an interface (in fact several) to add a bit to electronic memory. So, I'm not sure that the ability to act "outside granted abilities" is necessarily an indicator of ability to reason. "Decision making" it can certainly do, although the models I've tested seem to be specifically trained not to "make decisions." Again, iteration here gets the best result, but I tried this: "I'm currently a programmer with a stable income. Should I move to Hollywood and try to become an actor?" It returns a list of pros and cons, questions to ask, and then defers decision making. I then prompt, "Which would you do?" It very verbosely tells me why, "reasonably," retaining my proverbial current position would be better, but again defers the actual "decision." Then I just prompt, "Summarize your decision in one sentence." It replies, "I would likely maintain my programming career and explore alternative creative outlets that align with my skills and interests, prioritizing stability and leveraging transferable skills while still allowing for creative expression." We can even get better with, "Rephrase your summary as an absolute," yielding, "I will absolutely maintain my programming career and explore alternative creative outlets that align with my skills and interests." That's a decision. Because it's been clearly trained to avoid "deciding" things for people, it takes a couple steps to get that decision out of it, but it's fully capable. That's what the concept of the "agent" achieves, is automating that iteration to yield the desired outcome, with respect to the principle of "garbage in, garbage out." Now, because it truly has no preference, I can turn around and prompt, "Using deductive logic, explain why you would move to Hollywood and pursue an acting career instead of keeping your current job," and it will happily change its mind at my behest with a rather convincing argument with "appropriate" premises and conclusions about the value of creativity over stability. It's all about providing the proper context to make the "right" decision. Technically, a self-driving car meets all the rest of your requirements - and although it's not an LLM, it could certainly be interfaced with one, quite easily I suspect. They do all of the "correctly constructing future scenarios, [and] also scoring them with respect to expected rewards, and taking into account [possible future] actions of other actors." I can easily see fleets of driverless taxicabs within a very short time, with only a few people around to plug in/refuel the fleet. Airplanes have been fully capable of and utilizing automatic flight from ground to ground for quite some time now - the rare incident is typically caused by human or mechanical malfunction. GPT customer support models often score higher in customer satisfaction ratings than their human counterparts. This generation of "AI" already has us questioning the definitions of certain terms and phrases, and I think the next generation is going to force us to sharpen and refocus our definitions. I think any current arguments about "reason" and "decision making" at this point are purely semantics.
@@x82hammer28x You started out right, the model is only able to conjure something which resembels the context based on its training data. But then you jump to conclusions that because the responses seem similar to what humans would produce for such prompts, and because humans struggle with some prompts , then voila, the model is as good in "reasoning" as humans and we are just splitting hairs, and the "deliberation" can be done by just throwing in the previous output into a second round of generation. But what you describe is all that was already covered by Searle's "Chinese room" experiment. Having a humoungous lookup table in which you can do fuzzy lookups of any prompt does not create a capable reasoner because actual reasoning involves intentionality and iterative multi-step planning, it requires an execution of an iterative algorithm for calculating exepected penalties and rewards while keeping your (often contradictory) goals "in mind" during this evaluation, and only THEN producing the results (and possibly "changing your mind" multiple times in the process, while noticing mistakes etc.) Whereas the behavior of your LLM when it makes a mistake is to reinforce this mistake because it was fed into the context. We all remember how LLMs can go into endless loops repeating the same crap or arguing in favor of absurd propositions or even attributing these hallucinatios to their "interlocutor". This happens precisely because they lack this evaluative/planning/reflection capability, and it is up to you to prove that it can be achieved by just making a loop (and if you are not just making a loop, but instead very carefully selecting what you put into context to "guide" the model to arrive at the right answer, then YOU are doing the reasoning for the model, proving my point).
@@x82hammer28x Another nice example of models being dumb imitators rather than reasoners is code generation. The kinds of errors that they make: like outputting a plus instead of a minus.. because these operators happen to be "similar" by the distance metric in the embeddings. Similar to LLMs confusing "small" with "big" and the like. These are not errors that you would make if you, like a human, had a capability to introspect, self-evaluiate and follow goals, like a capable human reasoner does. It is reminiscent of young kids in school trying to answer "something" in hope of guessing correctly and pleasing the teacher (except that the intention of "cheating", which could be ascribed as a planning/reasoning trait in a child, is missing because the LLM has no intentions whatsoever). This is a general failure/limitation of the primitive reproductive algorithm (which has to be primitive to perform with acceptable speed), not something which you can patch up by throwing more training data into it.
"Getting the right prompts that match the LLM" yeah right because the amazing "artificial intelligence" is not even able to translate one format into another. Which tells you a lot about how intelligent it is.
The LLMs available to the public are not AGI. People aren't claiming they are.
Humans do not speak every language. Are they any less intelligent because (most) people can't execute all tasks given to them in English and Mandarin?
@@quinnherden Well, yes, they are translators. They should be able to translate the prompts at least but fail even at that. So shall we say, not even translators, but rather fuzzy query engines that demand query to be in a certain syntax. Less and less intelligent the more you look into it.
You could have saved everyone an hour and just said "don't use crewai". It's not a good framework, completely overhyped.
So was that the TLDW in the end then as I can't find that part?
The stock footage is unnecessary and lowering the quality of your content. Get rid of it. It's more interesting to look at an empty Colab notebook while you are talking rather than shitty stock footage (same as used by the other "AI boys").
slow n boring
I don’t understand what course you’re trying to sell me
This was great. I understand this stuff already, and this is a great, thorough overview
for me, it was really helpful to get your views about the other frameworks, like your comment about langgraph. clearly, there are more than one option, and knowing which one to focus one's energy is really helpful.
Totally agree with you on the importance of figuring out which agent framework to bet on / up-skill oneself. For me its between Crewai ai (powered by Langchain) vs Autogen (built & backed by Microsoft AI research team) With Microsoft’s backing, maybe Autogen will be the winner? (in-terms of being the de-fecato agent framework)