Awesome video! By the way, in general your format, pacing, conciseness and the cool but also utilitarian "code floating above your speaking" is a nice touch! I know making these videos takes a lot so just want to give kudos where it's clearly deserved! There's so many "faceless", auto-gen'd content or just click-baity low-content videos so this was a welcomed break from that!
I used CrewAI one or two times. It's simple but takes a lot of time, a lot of issues, a lot of tokens, and came to the same conclusion as yours! By the way, actually, after a lot of "improvements", it becames confusion instead of the simplicity of the beginning. Today I use light weight and straight forward automations!
loved seeing this Matt. I have spent far too much time trying to write code that fits into one agentic framework or another, whilst always feeling it should be the other way around. This is a real gem of a video and anyone that says you're an idiot, just hasn't used any of the available agent frameworks. The community was crying out for this video! Bravo sir.
I am using LangGraph which is also very deterministic. It uses a state and you can define functions as nodes (agents) thus for some tasks I avoid LLM which speeds up the process… The downside is of course that it takes some time to learn it. But if you do write a couple of them you can use them as templates for new projects. CrewAI is inconsistent and not useful for production at least in my case…
This video inspired me to use xstate + ollama json/functions instead of langgraph or crewai. I have made so much more progress with my advanced AI agent workflows and learned quite a bit more about state and complex logic in the process. Thanks for this video.
all frameworks are hype. the time it takes to learn their documentation/APIs is better spent implementing the custom solutions your project needs, and you won't be held back.
I have to admit that I've had more success just using the OpenAI API than from my experiments with CrewAI, or AutoGen, where I often feel like I'm fighting the framework. Good video
I used CrewAI for an internal prototype and my main issue is the framework has near zero observability. If I put crews into FastAPI async, guess what? They just fail without any notification at all. Plus the only TH-cam video claiming to use FastAPI with crew that I could find seems fake. The guy doesn’t even show his code, probably because it didn’t actually work and he had to hardcode his app to get the video out. There’s a lot of money sloshing around for agentic AI and a lot of short cuts are being taken. In the last month Crew has rewritten its code base and examples multiple times. I think AI is helping as it’s quickly becoming a useless dumpster fire. I can’t use Crew in a real application. But it’s certainly great for TH-cam videos and paid promotional marketing! 🎉 Thanks so so much for making this. You’ve made my life much easier by pointing out simpler examples.
It makes a lot of sense actually. If anyone is a programmer I’m sure one can use the logic and write functions to automate it. For normal humans, we can only try to come up and refine the prompt to do a task, then take that output and feed it to another agent and so on.
True... But trust me, take the time to learn. I'm still mad at myself for wasting two years from 1996 to 1998 with Microsoft Front Page when I could have just learned HTML in 3 months.
@@johngoad why would you be mad at Front Page? lol That was a good program back then. If you started doing web at that time those two years at least helped you with content structure.
I've used CrewAI pretty extensively. It is good for a specific type of agentic app, namely agents executing tightly structured processes in a coworking environment, but it is inflexible and rather buggy, especially if you're not using OpenAI's LLMs.
I love what you have been sharing on this channel. As a developer I really enjoy the content like in this video which is not just a tutorial on a tool but knowledge that is deeper. Also less code and more control, yes please !
Great video Matt! I fully agree with what you say. I've been on this track for a long time and have built a TS framework for it. I could not understand why all these agentic tools were so complex (and required agents to make obvious decisions which they sometimes failed at) and why everyone was hyping them so much (especially since you can make much more reliable tools in a simpler way). At the same time, I was unable to give words to why I felt this way. Your video helped me give words to why I felt this way! Thanks!
You are spot on and shows the difference between "letting the ai do the work" vs "making a good, resource efficient, lean process". This is why softeare engineers will have jobs for a while- along with the issue of mission critical reliability and confidence.
Thank you, Matt. You're right. If the process is simple and known, no agent is needed. Agents are meant for harder problems, e.g. independently figuring out the process to achieve a goal.
So true Matt. I’ve been trying to achieve code research tasks using a RAG and Ollama with local models and CrewAI framework for about two months, with no good results so far. Where a simple chat with Ollama and feeding the chat with results from the same RAG gives way better results and consistency. I’m keeping the CrewAI program just for evaluating new updates and see if it’s improving over the time with same agents and prompts…
It makes a lot of sense. As someone who is only just getting started playing with local LLMs, something like CrewAI seems easiest. Maybe what we really need is some good tutorials for leveraging existing libraries and techniques like you did. Thank you for making this - looking forward to your next videos!
Tools like crew have amazing intentions but it still has a bit of work to do to become something useful. It’s already starting with a lot of baggage but hopefully with the money they raised they can get some of that solved and be something very cool. M
Completely agree, before I start to learn Python, I just implement the things I really need with the language I already know. Agents is just an architectural pattern, you don't need a framework.
Function calling⁉️😂 there’s got to be a better name… I’ve built a few dozen personal projects in CrewAI and while I enjoy working with the framework, I agree it’s not the right tool for every agentic workflow. It is hard to debug currently. I really like your content, brings some fresh air into the space and the motion graphics are sick! 💯
Thank GOD! One person I can always trust. I was exactly searching for this. :) I really dint want to use any of the "decorators" that mascarade as a framework :P
If i can suggest something for the ranking formula, use coefficients from geometric series. 1/2^n where n is the position of the term, this will create no overlaps between ranked videos in the ranking space. P.s. thank your for this video.
One question. Incase I want to have two tasks that I want the ai to choose from. How do I do that? will it be a new ollama.generate , to choose and then an if else on top of that ?
I've tried a few different ways of getting more complex tasks done. I like the simplicity of this one. It fits my experience in that I find unless I plan on use a framework a lot, it's easier to just program the task(s) in a straight forward manner rather than take the time to learn a new package. I recall learning to use a MVC package to manage a somewhat complex web site. It really took longer than if I had just written it from scratch. I work on my own and seldom need to share my code with anyone else. Not that I don't want to but it never seems to work out.
Spot on. Agent frameworks are good for chatbots that need to deal with a large range of tasks...but well defined workflows can just be LLM-based automations with a human in the loop
Matt, new to agentic flows I’ll admit, but I was wondering if a n8n workflow could be the “right tool for the job” (where the flow can be defined)? 4:34
Thanks ... I appreciate you POV... I have used CrewAI, AutoGen and more.. I have wasted tons of time learning stuff that I don't use because at some point (usually 3/4 of the way in) I realize the limitations. I mostly just learn why it is not for me. Not sure if I can put a finger exactly on what it is, but it seems to me like most stuff is built these days with training wheels.
I was fixated on using CrewAI and langchain for agent building. Thanks for sharing your perspective. It's more clear to me on when exactly to use these frameworks.
anyway, my idea of a good agent framework would include this: in a userfriendly webinterface i can write down the request to find a good solution for an idea or project and the agent framework would come up with the choice of the appropriated agents and they identify all the tasks they need to get done to achieve the goal.
Love the awkward silence. As a not-coded-forever person, who roughly understands what you are doing vs crew AI…. I like your approach. Simpler is generally better and faster… BUT may not be an extensible in the future. Which may be irrelevant in this scenario.
Hey! Thanks for the video. But i am just wondering when would be a use case for agentic workflow? I understand that your solution is proposing a prompt flow instead
Yes, i am asking for a usecase for agents. Andrew Ng once said that you can improve LLM performance by introducing agentic workflow. Maybe I misunderstand. Your implementation to me looks more like a prompt flow (a pipeline of prompts) which does not have agents. Is that correct?
An agent is nothing more than a prompt to a model and then sometimes a loop to try to improve. None of the frameworks offer the ability to do that without way too much overhead. So it’s easier to write and debug a simpler flow that can do all of that. I think you are choosing a narrow definition of an agent.
Yep, this was what I was thinking too. LangGraph, crewAI etc. are cool to get started and build your first agent, but they really have a lot of unnecessary components, and they're creating complicated abstractions over pretty simple components
interesting, maybe a nocode/lowcode option would be a better alternative, especially for non programmers? Do you have preferences for low code tools for agents? Cheers
Tools/function calling is the proper way to do this but it becomes expensive very quickly because of the additional tokens used on every call. The frontier providers are all currently workng on tackling this and hope tools is in scope. Autogen is probably the most developed tool right now but I've found it to be problematic as the AI's can get stuck in loops !
I made a few rag thing in crewai, and while I love it, there's some huge gaps. Like you mentioned, when you know a tool, just tell it to use a tool. I find it cumbersome not to be able to properly debug prompts without using agentops or something. I do actually like the agents vs tasks vs tools solution, but in practice, the agent almost becomes an LLM connection rather than it's own thing.
I think it is a matter of using abstractions. If you just need to get things done - it is ok to use simple tools but when your prototype becomes more and more complicated you may need more abstractions in order to maintain all this stuff (if you use langchain for simple RAG - it will be easier to add extra LLM type support). I do like crew and langgraph because it is easier to maintain pipelines when you need to check agent’s results automatically, to make some decisions based on that or to have cycles. I prefer to use LangChain even just for simple LLM calls because it is easier to switch to another LLM and not to deal with issues like “Claude has system prompt as another parameter instead of first chat message” etc
While I agree with the sentiment that simpler is better. Sometimes you don't know that you could benefit from the ability to let the LLM talk back and forth between other LLMs until you've tried it on several workflows. If your plan is super well defined then it makes sense to simply the design, but for now I'll continue using crewai to scaffold out AI based workflows till I can determine the true limits of the models that are available. Good video though! Thank you for putting it together. It made me think!
you're right, but your approach require a deeper knoledge of programming, and all that tools are moving on the hopposite direction. As you told, the right tool depend of the thing you're doing, but I think the more complex the thing will be, less advantage your solution will get. last but not least, business side, people that know that tools ( CrewAI and the others) will have a vantage.
The way I see it, with this workflow, is feels like you're actually creating something, and that you know every piece of the puzzle. As opposed to relying on some framework's internal workings, of which you have to spend time learning and adapting your way of reasoning to that of the authors'. Of course, your point still stands. This is only valid for programmers, not normal users.
Crew and the others are dev tools. My approach requires less knowledge and gets it done. Those other tools simply complicate things while adding nothing of value.
Coding is a pretty basic skill anyone can learn. Not much different from speaking a language. That said there are plenty of ways to do this stuff without coding too
Why do you think there are a good number of codebases going to GO over Python? I noticed Fabric Framework did this a little while ago. Also, love your videos subscribed for sure.
Packaging is a huge one. Distributing any product that requires Python means a lot of support issues. And Python tends to be slower. I’m not sure I see a reason any package would choose to start with Python. There really isn’t any benefit.
@@technovangelist, thank you for the response! As a full-stack developer, I always use Python on the backend, and I would love to learn Go and give it a shot for upcoming projects. I've used Python for so long that I never learned Go. I would like to try it out and see how the performance and code differ. I just found your channel yesterday and love it.
what if i cant code? (just learning - got my computer yesterday!) - I once coded in basic, html, asp - but that was MANY years ago - so basically a novice again. I have a clear app that I want to build for the OHS industry (just one, very specific area obviously). Do I have to learn python or GO or something else just to build my app? Would I be better off outsourcing?
I can see a toolbar under your TH-cam video on your screen containing timestamped notes, AI notes, summarize video, and screenshot. Can you please share what is it.
So, is the idea here to make some very simple functions or methods for the LLM to call to get the job done, then use function calling to get the LLM to do the job?
not really. the whole thing is about automating your processes with agents, except the agent frameworks are garbage. function calling does not refer to models calling functions.
@@technovangelist Yes, unfortunate that every time we talk about Function Calling, we have to explain that the LLM does not actually call any functions. :(
Wish there was a channel like this for those of us who don't code but want to get the same things out of AI, which. is really where AI should be going. We shouldn't care of it's Python, Go, Rust etc. We should just describe what we want and the finished product, with some iterations, should pop out the other end.
Thanks for the video! I agree the frameworks are not ready yet; SuperAGI is another example. Very, very promising but not there yet. But most people just don't know how to code so the frameworks like CrewAI are simply the only choice. Maybe in a next video you could show how the results of a solution with coding are faring against CrewAI & co. Would give a perspective on the maturity status of these frameworks.
What's the tldr on what is worth the time today to learn to make agents? (For someone who understands coding, a bit, but is not an active programmer*?)
At 6:50 for the formula, I think the polynomial needs "+" (instead of the "x") after the numerics so you have separate terms. Oh actually at 9:12 looks like you get 'em, okay now..
so you moved from an agentic approach to a series of ai calls with some tooling. ok - sometimes this might be useful. to me, the concept behind crewai, langgraph and others points more towards having agents (ais with - sometimes multiple - tools/functions) organize themselves to generate/create some output in a non procedural manner.
No. If you want to be lazy, sure let the model guess the process and it gets it right some of the time. Or just put in the work and get it done. The frameworks don't really add anything. Use the AI and the tools. Organizing themselves is a gimmick
6:40 I'm confused about this equation. Score = (Views/Subs)*.4 * (Likes/Views)*.3 * (Comments/Views) * .2 * (1/DaysSincePub) The constants can be multiplied together, and two of the Views cancel out. So I think you end up with: Score = 0.024 * (Likes * Comments) / (Subs * Views * DaysSincePub) Is there any particular reason it was formatted the way shown at 6:40? "The views are more important than the likes, and the likes are more important than comments." I'm not sure if the formula is capturing what the wording here implies. Since Views ends up in the denominator, views actually end up a bad thing. Higher views = lower Score. Is that intended...? Edit: OK, nevermind. Looks like at 9:10, these values are added, and not multiplied. Which means they don't cancel out. So I guess it's fine :)
Thanks Matt, I can't get it to run, but thanks. I guess you're using linux, and I am on windows. I am not a programer and all the stuff I have used works with node.js not bun? Not sure what bun or bun.js is, I tried and could not get this to run. But thanks for taking the time to share.
Great points about Kubernetes being overkill and just the over focus on Python etc. I am building an open source tool inspired by Laravel PHP just to help me see how it is not the language but so much more. And good points about automation, to me I want normal non coding users to see the power and ease of this as they see the power of prompts. Anyways great one!
Hi Matt, thanks for this video. But what would you recommend for the following? I want to develop software with agents. The target is not a specific workflow but a actual software that they need to create.
Great one, and fully aligned with you. People are always focusing on the tool instead of focusing on the need. Most of the agentic framework I’ve seen so far were falling apart as soon as you’re trying to implement something a little bit complex … and rely and local models. And the one big problem they all have (but somehow your code as well) is that they are as good as your prompting skills are … which on top will vary depending on the model you use …
the prompting issue I think is valid. But still, we don't really know what works best yet, it's no man's land. I think that overtime you could try out different prompts and see how they perform, but obviously it takes time to collect this data.
Hey Matt, would you consider a follow-up to this using one of crewAI's actual examples? I think that'd really help to illustrate the differences between the two approaches. They list them all on their docs page (not sure if I can link here)
Why did ollama abandoned python? I thought python runs the ai-show and is the best for it. I also see it everywhere as a prerequisite to build sold ai career. Can you elaborate on that? Thanks 👋 😎
There is nothing special about Python that makes it any better for ai than any other language. And considering how easy it is to screw up a Python environment there are definitely reasons to avoid it. For some specific scenarios there are some libraries that may exist on Python and not in other language but that’s not true for everything. Ollama was only written in Python for about 3-5 weeks and then we hit a wall of what we could do. It’s far easier to install ollama now that it is go based than before when it was Python based.
I think similar. But to be honest I haven't tried those frameworks. From what I saw they seem to be too complex for most of the tasks you will need an agent. Most of the time - my point of view - you just need a specialized agent/workflow and not an agent who "can" do everything. Otherwise it would get to complicated as Matt is emphasizing here. I was discovering "instructor", a python library for getting structured output from LLMs in JSON using OpenAIs client and for a bunch of other providers. Since a few days there is a new library called "ollama-instructor" with the same approach but using Ollamas python client natively. Both use Pydantic for creating and validating the JSON response from the LLM. Opinion: With "instructor" or "ollama-instructor" you can create specialized workflows where you have more control over the flow itself and the outcome as with an agent framework.
Was adding so much code to get Crew to work as I wanted til I got to the point where the loss of control not knowing the input prompts Crew was adding and looking at what I had already written I had to scratch my head wondering how much more was necessary to have my own agent system. Couple of lines of code later and I was free of Crew. Not looking back.
I appreciate the time you took to put this together. However, I am always amused when I click on a video that says I’m overcomplicating things, then proceeds to discuss complicated things. I’ve gotten to the point where I was able to get Ollama defaulted to Llama3 working locally with AnythingLLM on the front end, but I’m not a developer and it seems like it is necessary to be one to grasp some of these concepts.
Ahh well the video was about agent frameworks but if not a dev you can’t use any of those anyway, crew, autogen, and others are dev tools, as is ollama.
@@technovangelistYeah, I’m a hack. Basic Python, PHP, JS. At this point just happy I was able to work through the stack traces to get this all running locally. Sort of. Very slow but the PC is old. Had to install OpenAPI, CUDA, Visual Studio Code, updated Intel drivers…it has been an experience 😆
Very interesting. You brought up a really good point. I was actually wondering why I never really got into CrewAi, and the simple answer is that it was never really useful. Reliability and using as few tokens as possible are important. Actually you want to use LLMs as little as possible
Great video. Keeping with the simplicity topic. I often find myself say promoting chat gpt 5 times. I then want to simply run a bunch of data through those same 5 prompts (with the ability to see and adjust output between each prompt if desired). What’s the simplest way to do this in your opinion?
He reminds me of a anthropology teacher in high school that was super cool. He kept the subject matter interesting & focused. I like this channel. Well Done
NLU and LLM the way to go, NLU still king as is assured "CERTAINTY", intersplice with tactical LLM agents, LLM with Rag Educate an Agent GUI. This is the model that is forging ahead. If you want a real service there's about another 50 things you need to consider, and most tools, especially codey tools cover
I et the feeling that the companies that release AI are not wanting people like us to use them for automation, and aggressively push back against any "simple" data exchange vectors such as the command line and console. Like I am looking at console emulation for hosting a local ollama install to be able to exchange data, there is no "managed code" that seems to allow simple exchange or interfacing the AI.
No. Different companies release products for different audiences. The most powerful raw products are always going to be developer focused so they can build amazing products with it. Ollama is for developers first. That’s the target. So videos about it are usually focused around the code.
Awesome video! By the way, in general your format, pacing, conciseness and the cool but also utilitarian "code floating above your speaking" is a nice touch! I know making these videos takes a lot so just want to give kudos where it's clearly deserved! There's so many "faceless", auto-gen'd content or just click-baity low-content videos so this was a welcomed break from that!
I used CrewAI one or two times. It's simple but takes a lot of time, a lot of issues, a lot of tokens, and came to the same conclusion as yours! By the way, actually, after a lot of "improvements", it becames confusion instead of the simplicity of the beginning. Today I use light weight and straight forward automations!
use openrouter to use cheap models
You are spot on, Matt. CrewAI reflects all the new tools in the AI space in general: too much hype, not enough utility... yet.
loved seeing this Matt. I have spent far too much time trying to write code that fits into one agentic framework or another, whilst always feeling it should be the other way around. This is a real gem of a video and anyone that says you're an idiot, just hasn't used any of the available agent frameworks. The community was crying out for this video! Bravo sir.
I am using LangGraph which is also very deterministic. It uses a state and you can define functions as nodes (agents) thus for some tasks I avoid LLM which speeds up the process… The downside is of course that it takes some time to learn it. But if you do write a couple of them you can use them as templates for new projects. CrewAI is inconsistent and not useful for production at least in my case…
Did you use langchain 2.0?
You could also just get flowise and then everything can be done with a gui. It just uses nodes which means you dont need to look through the code.
@@pin65371 Flowise has the customizability to build an agency? I've been trying to decide which Framework
@@pin65371 cool I am using LangGraph too, but I thought I would love a gui flow, thanks❤️❤️❤️
Flowise's UI is buggy at the moment. Also they lack documentation.
This video inspired me to use xstate + ollama json/functions instead of langgraph or crewai. I have made so much more progress with my advanced AI agent workflows and learned quite a bit more about state and complex logic in the process. Thanks for this video.
all frameworks are hype. the time it takes to learn their documentation/APIs is better spent implementing the custom solutions your project needs, and you won't be held back.
Wise
@@petemoss3160 exactly
This is what we need. Crewai... too much hype and not enough utility. Thanks for sharing.
I think a BPMN 2.0 tool with service task is enough instead of a dedicated agent framework.
I have to admit that I've had more success just using the OpenAI API than from my experiments with CrewAI, or AutoGen, where I often feel like I'm fighting the framework.
Good video
Thanks for stopping by.
It is true for now... but then you have to pay and it is not private.
I used CrewAI for an internal prototype and my main issue is the framework has near zero observability. If I put crews into FastAPI async, guess what? They just fail without any notification at all. Plus the only TH-cam video claiming to use FastAPI with crew that I could find seems fake. The guy doesn’t even show his code, probably because it didn’t actually work and he had to hardcode his app to get the video out.
There’s a lot of money sloshing around for agentic AI and a lot of short cuts are being taken. In the last month Crew has rewritten its code base and examples multiple times. I think AI is helping as it’s quickly becoming a useless dumpster fire. I can’t use Crew in a real application. But it’s certainly great for TH-cam videos and paid promotional marketing! 🎉
Thanks so so much for making this. You’ve made my life much easier by pointing out simpler examples.
It makes a lot of sense actually. If anyone is a programmer I’m sure one can use the logic and write functions to automate it. For normal humans, we can only try to come up and refine the prompt to do a task, then take that output and feed it to another agent and so on.
True... But trust me, take the time to learn. I'm still mad at myself for wasting two years from 1996 to 1998 with Microsoft Front Page when I could have just learned HTML in 3 months.
@@johngoad why would you be mad at Front Page? lol That was a good program back then. If you started doing web at that time those two years at least helped you with content structure.
@@vitalis Back then, production level code required hand written HTML.... My company had 20 writing HTML and 5 programmers. True story...
I've used CrewAI pretty extensively. It is good for a specific type of agentic app, namely agents executing tightly structured processes in a coworking environment, but it is inflexible and rather buggy, especially if you're not using OpenAI's LLMs.
I love what you have been sharing on this channel. As a developer I really enjoy the content like in this video which is not just a tutorial on a tool but knowledge that is deeper. Also less code and more control, yes please !
Thanks for the comment and for being a member
Great video Matt! I fully agree with what you say. I've been on this track for a long time and have built a TS framework for it. I could not understand why all these agentic tools were so complex (and required agents to make obvious decisions which they sometimes failed at) and why everyone was hyping them so much (especially since you can make much more reliable tools in a simpler way). At the same time, I was unable to give words to why I felt this way. Your video helped me give words to why I felt this way! Thanks!
Simplicity is the product of wisdom, and that comes from effective experience. Most excellent video.
You're the best man! No one does anything as exciting as you do, with simplicity and exact methodology!
You are spot on and shows the difference between "letting the ai do the work" vs "making a good, resource efficient, lean process". This is why softeare engineers will have jobs for a while- along with the issue of mission critical reliability and confidence.
AI will slowly take jobs, but more senior devs are pretty safe for a while.
Thank you, Matt. You're right. If the process is simple and known, no agent is needed. Agents are meant for harder problems, e.g. independently figuring out the process to achieve a goal.
Agents are no more than a prompt. Sometimes a loop. But crew and similar tools are just a gimmick.
@@technovangelist haha....well said
I have yet to make a choice and commit to a framework. Glad I watched this video first.
So true Matt. I’ve been trying to achieve code research tasks using a RAG and Ollama with local models and CrewAI framework for about two months, with no good results so far. Where a simple chat with Ollama and feeding the chat with results from the same RAG gives way better results and consistency. I’m keeping the CrewAI program just for evaluating new updates and see if it’s improving over the time with same agents and prompts…
It makes a lot of sense. As someone who is only just getting started playing with local LLMs, something like CrewAI seems easiest. Maybe what we really need is some good tutorials for leveraging existing libraries and techniques like you did.
Thank you for making this - looking forward to your next videos!
Tools like crew have amazing intentions but it still has a bit of work to do to become something useful. It’s already starting with a lot of baggage but hopefully with the money they raised they can get some of that solved and be something very cool. M
My wife told me not to bother with these brand new frameworks (LangChain, CrewAI, Semantic Kernal, Amazon Q). Make my own. She was right.
Your wife sounds awesome.
By definition, the wife is ALWAYS right. So . . .
True
I love this. Exactly. Spot on. This is why I'm building my new project from scratch.
Can you talk about how you deploy this agent to run it periodically in the cloud?
I’m marking any emails I get with the word “delve” in them as spam 😂
Genius! 😆
Completely agree, before I start to learn Python, I just implement the things I really need with the language I already know.
Agents is just an architectural pattern, you don't need a framework.
Function calling⁉️😂 there’s got to be a better name…
I’ve built a few dozen personal projects in CrewAI and while I enjoy working with the framework, I agree it’s not the right tool for every agentic workflow. It is hard to debug currently.
I really like your content, brings some fresh air into the space and the motion graphics are sick! 💯
Thank GOD! One person I can always trust. I was exactly searching for this. :) I really dint want to use any of the "decorators" that mascarade as a framework :P
If i can suggest something for the ranking formula, use coefficients from geometric series. 1/2^n where n is the position of the term, this will create no overlaps between ranked videos in the ranking space.
P.s. thank your for this video.
One question. Incase I want to have two tasks that I want the ai to choose from. How do I do that? will it be a new ollama.generate , to choose and then an if else on top of that ?
I've tried a few different ways of getting more complex tasks done. I like the simplicity of this one. It fits my experience in that I find unless I plan on use a framework a lot, it's easier to just program the task(s) in a straight forward manner rather than take the time to learn a new package. I recall learning to use a MVC package to manage a somewhat complex web site. It really took longer than if I had just written it from scratch. I work on my own and seldom need to share my code with anyone else. Not that I don't want to but it never seems to work out.
Spot on. Agent frameworks are good for chatbots that need to deal with a large range of tasks...but well defined workflows can just be LLM-based automations with a human in the loop
Yep. And in most cases, if you think there isn't a well-defined workflow, it just means you haven't really thought about it, and there probably is.
Matt, new to agentic flows I’ll admit, but I was wondering if a n8n workflow could be the “right tool for the job” (where the flow can be defined)? 4:34
I am just getting started on learning about agents. So glad to find this video and hear your perspective.
Glad to be of help
This is the most sane thing I've seen about Crew AI. Thank you.
your welcome
Thanks ... I appreciate you POV... I have used CrewAI, AutoGen and more.. I have wasted tons of time learning stuff that I don't use because at some point (usually 3/4 of the way in) I realize the limitations. I mostly just learn why it is not for me.
Not sure if I can put a finger exactly on what it is, but it seems to me like most stuff is built these days with training wheels.
Excellent material Matt! thank you. I am delighted to watch your presentations. This is my favorite channel.
I was fixated on using CrewAI and langchain for agent building. Thanks for sharing your perspective. It's more clear to me on when exactly to use these frameworks.
Thank you for your video! I can't find the index.ts file. What do I need to run in the index.ts file? Thanks!
You mentioned Charles Petzold, when i first read it i was 18. A lot of time has passed.
is it possible for long range task with non deterministic input and output
Of course
anyway, my idea of a good agent framework would include this: in a userfriendly webinterface i can write down the request to find a good solution for an idea or project and the agent framework would come up with the choice of the appropriated agents and they identify all the tasks they need to get done to achieve the goal.
Love the awkward silence.
As a not-coded-forever person, who roughly understands what you are doing vs crew AI…. I like your approach. Simpler is generally better and faster… BUT may not be an extensible in the future. Which may be irrelevant in this scenario.
Hey! Thanks for the video. But i am just wondering when would be a use case for agentic workflow? I understand that your solution is proposing a prompt flow instead
Two phrases for the same thing
Or if you are asking when is the use case for these frameworks? Never seen one that justifies the extra work to complicate the flow.
Yes, i am asking for a usecase for agents. Andrew Ng once said that you can improve LLM performance by introducing agentic workflow.
Maybe I misunderstand. Your implementation to me looks more like a prompt flow (a pipeline of prompts) which does not have agents. Is that correct?
An agent is nothing more than a prompt to a model and then sometimes a loop to try to improve. None of the frameworks offer the ability to do that without way too much overhead. So it’s easier to write and debug a simpler flow that can do all of that. I think you are choosing a narrow definition of an agent.
Yep, this was what I was thinking too.
LangGraph, crewAI etc. are cool to get started and build your first agent, but they really have a lot of unnecessary components, and they're creating complicated abstractions over pretty simple components
Great Stuff! totally agree, crewai main idea behind is a - HYPE, not code, not devs, not user experience. so good to know i’m not alone here 🙌😀
Yes, please share your code for this and other practical use cases.
What do you think about agent zero?
interesting, maybe a nocode/lowcode option would be a better alternative, especially for non programmers? Do you have preferences for low code tools for agents? Cheers
I absolutely love Matts Videos, such a cool person, and keeps things simple.
Tools/function calling is the proper way to do this but it becomes expensive very quickly because of the additional tokens used on every call.
The frontier providers are all currently workng on tackling this and hope tools is in scope.
Autogen is probably the most developed tool right now but I've found it to be problematic as the AI's can get stuck in loops !
There are no additional tokens used for function calling in ollama. Not sure what you mean. And there is no cost beyond you machine and gpu.
Apologies I'm referring to running hosted models operating at scale as opposed to running models locally.
When Matt talks.. I listen! Spot-on dear Sir
I made a few rag thing in crewai, and while I love it, there's some huge gaps. Like you mentioned, when you know a tool, just tell it to use a tool. I find it cumbersome not to be able to properly debug prompts without using agentops or something. I do actually like the agents vs tasks vs tools solution, but in practice, the agent almost becomes an LLM connection rather than it's own thing.
I think it is a matter of using abstractions. If you just need to get things done - it is ok to use simple tools but when your prototype becomes more and more complicated you may need more abstractions in order to maintain all this stuff (if you use langchain for simple RAG - it will be easier to add extra LLM type support).
I do like crew and langgraph because it is easier to maintain pipelines when you need to check agent’s results automatically, to make some decisions based on that or to have cycles.
I prefer to use LangChain even just for simple LLM calls because it is easier to switch to another LLM and not to deal with issues like “Claude has system prompt as another parameter instead of first chat message” etc
As the project gets more complicated it’s even more important to keep it maintainable and get rid of these tools that complicate things like crew.
Spot on. Amazing how clear you explain things. love this channel. kudos
While I agree with the sentiment that simpler is better. Sometimes you don't know that you could benefit from the ability to let the LLM talk back and forth between other LLMs until you've tried it on several workflows. If your plan is super well defined then it makes sense to simply the design, but for now I'll continue using crewai to scaffold out AI based workflows till I can determine the true limits of the models that are available. Good video though! Thank you for putting it together. It made me think!
but you don't need crew to add that. crew adds nothing and just complicates
you're right, but your approach require a deeper knoledge of programming, and all that tools are moving on the hopposite direction. As you told, the right tool depend of the thing you're doing, but I think the more complex the thing will be, less advantage your solution will get. last but not least, business side, people that know that tools ( CrewAI and the others) will have a vantage.
The way I see it, with this workflow, is feels like you're actually creating something, and that you know every piece of the puzzle. As opposed to relying on some framework's internal workings, of which you have to spend time learning and adapting your way of reasoning to that of the authors'. Of course, your point still stands. This is only valid for programmers, not normal users.
Crew and the others are dev tools. My approach requires less knowledge and gets it done. Those other tools simply complicate things while adding nothing of value.
@@technovangelist Is there a way non-devs can get their processes automated without having to learn coding?
Coding is a pretty basic skill anyone can learn. Not much different from speaking a language. That said there are plenty of ways to do this stuff without coding too
Finally, someone making sense :) I would never use crewai it's just hard work and in production its gives more problems than it solves.
Thank you great help in this jungle. I wonder what u think of the most recent AI dial odyssey epam framework. Its much bigger than crewAI… saluti
Why do you think there are a good number of codebases going to GO over Python? I noticed Fabric Framework did this a little while ago. Also, love your videos subscribed for sure.
Packaging is a huge one. Distributing any product that requires Python means a lot of support issues. And Python tends to be slower. I’m not sure I see a reason any package would choose to start with Python. There really isn’t any benefit.
@@technovangelist, thank you for the response! As a full-stack developer, I always use Python on the backend, and I would love to learn Go and give it a shot for upcoming projects. I've used Python for so long that I never learned Go. I would like to try it out and see how the performance and code differ. I just found your channel yesterday and love it.
Hi Matt, would you consider making a video using XState v5 to manage the flow?
Did he cover he to pick the best “tool” for the job?
what if i cant code? (just learning - got my computer yesterday!) - I once coded in basic, html, asp - but that was MANY years ago - so basically a novice again. I have a clear app that I want to build for the OHS industry (just one, very specific area obviously). Do I have to learn python or GO or something else just to build my app? Would I be better off outsourcing?
You can learn how to code. You can even use LLMs to help you learn how to code...
Love the pragmatic approach, thank you for the video. Rethinking my LlamaIndex learning now.
I can see a toolbar under your TH-cam video on your screen containing timestamped notes, AI notes, summarize video, and screenshot. Can you please share what is it.
Memex
So, is the idea here to make some very simple functions or methods for the LLM to call to get the job done, then use function calling to get the LLM to do the job?
Yeah I think it's akin to saying "don't create a React/Vue/Angular/whatever whole SPA when you only need one simple landing page"
not really. the whole thing is about automating your processes with agents, except the agent frameworks are garbage. function calling does not refer to models calling functions.
@@technovangelist Yes, unfortunate that every time we talk about Function Calling, we have to explain that the LLM does not actually call any functions. :(
Wish there was a channel like this for those of us who don't code but want to get the same things out of AI, which. is really where AI should be going. We shouldn't care of it's Python, Go, Rust etc. We should just describe what we want and the finished product, with some iterations, should pop out the other end.
That’s probably a few months to a year out
Try Flowise, it is not perfect but can help get you going.
"Right, I don't code, but I want to be able to do the same as people who can code do, nevermind learning to code, that's for nerds."
Thanks for the video! I agree the frameworks are not ready yet; SuperAGI is another example. Very, very promising but not there yet.
But most people just don't know how to code so the frameworks like CrewAI are simply the only choice.
Maybe in a next video you could show how the results of a solution with coding are faring against CrewAI & co.
Would give a perspective on the maturity status of these frameworks.
Folks that don’t know how to code can’t use crew either
We need an LLM to build the agents and select the appropriate model for each task.
Not really. This stuff is pretty easy
@@technovangelisthow are you?
What's the tldr on what is worth the time today to learn to make agents? (For someone who understands coding, a bit, but is not an active programmer*?)
Just use the ollama api
At 6:50 for the formula, I think the polynomial needs "+" (instead of the "x") after the numerics so you have separate terms. Oh actually at 9:12 looks like you get 'em, okay now..
so you moved from an agentic approach to a series of ai calls with some tooling.
ok - sometimes this might be useful.
to me, the concept behind crewai, langgraph and others points more towards having agents (ais with - sometimes multiple - tools/functions) organize themselves to generate/create some output in a non procedural manner.
No. If you want to be lazy, sure let the model guess the process and it gets it right some of the time. Or just put in the work and get it done. The frameworks don't really add anything. Use the AI and the tools. Organizing themselves is a gimmick
6:40 I'm confused about this equation.
Score = (Views/Subs)*.4 * (Likes/Views)*.3 * (Comments/Views) * .2 * (1/DaysSincePub)
The constants can be multiplied together, and two of the Views cancel out.
So I think you end up with:
Score = 0.024 * (Likes * Comments) / (Subs * Views * DaysSincePub)
Is there any particular reason it was formatted the way shown at 6:40?
"The views are more important than the likes, and the likes are more important than comments."
I'm not sure if the formula is capturing what the wording here implies.
Since Views ends up in the denominator, views actually end up a bad thing. Higher views = lower Score.
Is that intended...?
Edit: OK, nevermind. Looks like at 9:10, these values are added, and not multiplied. Which means they don't cancel out. So I guess it's fine :)
Thank you so much for this video. Super helpful and well-presented. Your time and effort is most appreciated!
hi Matt what is your take on LangGraph ?
Your videos are always a great reality check. Thank you so much, Matt.
Glad you like them!
Thanks Matt, I can't get it to run, but thanks. I guess you're using linux, and I am on windows. I am not a programer and all the stuff I have used works with node.js not bun? Not sure what bun or bun.js is, I tried and could not get this to run. But thanks for taking the time to share.
Outstanding video. As a complete newbie this was intriguing, well explained and engaging.
New subber and bell set
Great points about Kubernetes being overkill and just the over focus on Python etc.
I am building an open source tool inspired by Laravel PHP just to help me see how it is not the language but so much more.
And good points about automation, to me I want normal non coding users to see the power and ease of this as they see the power of prompts.
Anyways great one!
Hi Matt, thanks for this video. But what would you recommend for the following? I want to develop software with agents. The target is not a specific workflow but a actual software that they need to create.
Not sure....all i can say is that I wouldn't recommend most of the existing frameworks
@@technovangelist oki, thanks for getting back at me.
Great one, and fully aligned with you. People are always focusing on the tool instead of focusing on the need. Most of the agentic framework I’ve seen so far were falling apart as soon as you’re trying to implement something a little bit complex … and rely and local models. And the one big problem they all have (but somehow your code as well) is that they are as good as your prompting skills are … which on top will vary depending on the model you use …
the prompting issue I think is valid. But still, we don't really know what works best yet, it's no man's land.
I think that overtime you could try out different prompts and see how they perform, but obviously it takes time to collect this data.
Relying on local models is bad? Better give all your stuff to openai, you're right
Hey Matt, would you consider a follow-up to this using one of crewAI's actual examples? I think that'd really help to illustrate the differences between the two approaches. They list them all on their docs page (not sure if I can link here)
The example I showed is very comparable to
@@technovangelist Absolutely. Apologies if I made it sound otherwise. Just being greedy since your content is the best
This is exactly why I switched to Yacana! CrewAI was a mess and wasn't able to call my functions. And also too much boiler plate code like you said.
Why did ollama abandoned python? I thought python runs the ai-show and is the best for it. I also see it everywhere as a prerequisite to build sold ai career. Can you elaborate on that? Thanks 👋 😎
There is nothing special about Python that makes it any better for ai than any other language. And considering how easy it is to screw up a Python environment there are definitely reasons to avoid it. For some specific scenarios there are some libraries that may exist on Python and not in other language but that’s not true for everything. Ollama was only written in Python for about 3-5 weeks and then we hit a wall of what we could do. It’s far easier to install ollama now that it is go based than before when it was Python based.
Your “awkward pauses” at the end remind me so much of Craig Ferguson’s late night shows.
Great timing. I was just about to dive into some of these, but I'm not a Python guy and I like your approach better. Good points re: utility. Thanks!
I think similar. But to be honest I haven't tried those frameworks. From what I saw they seem to be too complex for most of the tasks you will need an agent. Most of the time - my point of view - you just need a specialized agent/workflow and not an agent who "can" do everything. Otherwise it would get to complicated as Matt is emphasizing here.
I was discovering "instructor", a python library for getting structured output from LLMs in JSON using OpenAIs client and for a bunch of other providers. Since a few days there is a new library called "ollama-instructor" with the same approach but using Ollamas python client natively. Both use Pydantic for creating and validating the JSON response from the LLM.
Opinion: With "instructor" or "ollama-instructor" you can create specialized workflows where you have more control over the flow itself and the outcome as with an agent framework.
Was adding so much code to get Crew to work as I wanted til I got to the point where the loss of control not knowing the input prompts Crew was adding and looking at what I had already written I had to scratch my head wondering how much more was necessary to have my own agent system. Couple of lines of code later and I was free of Crew. Not looking back.
Well, what alternative do you use? You never showed us the app
I did. The alternative is to not use a framework since it doesn’t add any benefit.
@@technovangelist Sorry I am not that into tech, but I am trying to be. But what is the app you are using typing in the code?
Any editor will work. Lots of folks use VSCode. But it's probably best to learn how to use Python or Javascript first.
the silence part is so cool!
I appreciate the time you took to put this together. However, I am always amused when I click on a video that says I’m overcomplicating things, then proceeds to discuss complicated things. I’ve gotten to the point where I was able to get Ollama defaulted to Llama3 working locally with AnythingLLM on the front end, but I’m not a developer and it seems like it is necessary to be one to grasp some of these concepts.
Ahh well the video was about agent frameworks but if not a dev you can’t use any of those anyway, crew, autogen, and others are dev tools, as is ollama.
@@technovangelistYeah, I’m a hack. Basic Python, PHP, JS. At this point just happy I was able to work through the stack traces to get this all running locally. Sort of. Very slow but the PC is old. Had to install OpenAPI, CUDA, Visual Studio Code, updated Intel drivers…it has been an experience 😆
Very interesting. You brought up a really good point. I was actually wondering why I never really got into CrewAi, and the simple answer is that it was never really useful. Reliability and using as few tokens as possible are important. Actually you want to use LLMs as little as possible
I think it will mature and become a really great framework, but it's got some growing up to do.
Even better to not use any LLM at all, problem solved!
as always, love it! good job man !
Great video, efficient and effective.
moving your head Up and Down to stop your camera?! Wow, Great job man that is a pure epic script^^
Huh?
@@technovangelist now you got to do it for real lol - The People have spoken
what does this mean?
Great video. Keeping with the simplicity topic. I often find myself say promoting chat gpt 5 times. I then want to simply run a bunch of data through those same 5 prompts (with the ability to see and adjust output between each prompt if desired). What’s the simplest way to do this in your opinion?
He reminds me of a anthropology teacher in high school that was super cool. He kept the subject matter interesting & focused.
I like this channel. Well Done
Awesome, thoughtful video, Matt! And you have an amazing voice--you should license it to ElevenLabs. 😁
NLU and LLM the way to go, NLU still king as is assured "CERTAINTY", intersplice with tactical LLM agents, LLM with Rag Educate an Agent GUI. This is the model that is forging ahead. If you want a real service there's about another 50 things you need to consider, and most tools, especially codey tools cover
I et the feeling that the companies that release AI are not wanting people like us to use them for automation, and aggressively push back against any "simple" data exchange vectors such as the command line and console. Like I am looking at console emulation for hosting a local ollama install to be able to exchange data, there is no "managed code" that seems to allow simple exchange or interfacing the AI.
No. Different companies release products for different audiences. The most powerful raw products are always going to be developer focused so they can build amazing products with it. Ollama is for developers first. That’s the target. So videos about it are usually focused around the code.