wow. After playing with CrewAi, Langchain and Autogen... this blew my mind. Amazing work man. Please continue adding functionalities, you're on to something massive.
Best content. Real, no fluff! Thank you so much for not just making the videos but sharing your work! You are changing the world. My world for sure. Much love!
Okay why not add another agent that takes screenshots and uses gpt 4 vision with some saved images of great landing page designs to communicate with the developer to to create a much better design
Awesome. Would be really helpful if you could show a step by step for how to get set up from the very beginning. I got stuck on installation on Windows and can't get anything up and running
You can import them from the other agent's folder using standard python imports. However, if there is a need for this, most likely you are not splitting your agent roles right.
You don't understand what about is this game changing stuff. It's not about how quickly you create website like this, it's about finding new method. In few years it will be autonomous agency. But someone has to start developing. So your comment only show your old mindset. You live in old world.
Very cool…have you thought about somehow getting it to run inside a container? While I’m technical there’s no way I can adjust the code files…as you have…as not a pure dev….so won’t be testing this unless it’s on a dedicated machine…
@@vrsen This is a huge flaw to Agency swarms. I use Crew AI with Groq which delivers near instant responses with the speed being at real time I am able to to have my agents to talk to me in real sounding human like voices using text to speech platforms. Yet to experience a single hallucination.
how much did it cost in api costs to run the ai agent swarm and build the website shown in this video? In next video, can you also please show how to use gpt 3.5 to save costs on api?
In your opinion, are openai assistants better than langchain's custom agents? The impression I have is that the assistants already had all the tools ready while the agents would have to choose and manage vector database, retrieval and memory resources.
Yes, there are many reasons why assistants are better. The biggest one is function calling. It makes your agents a lot more reliable in production. Creating tools from scratch is not a big issue, especially with genesis agency. It also gives you a lot more control
How about a system to streamline lead generation and qualification, aiming to replace the traditional Sales Development Representatives (SDRs) role? Researching leads can be time-consuming, so automating this process could provide a solution by delivering pre-qualified leads ready for outreach.
I have no idea why, but I haven't been able to get your code working, I keep trying as it's massively cool. Just trying to run the `agency-swarm genesis` command, I get: usage: agency-swarm [-h] {create-agent-template} ... agency-swarm: error: argument create_template: invalid choice: 'genesis' (choose from 'create-agent-template') That's after having pip uninstall agency-swarm, then pip install agency-swarm, to ensure I'm running the latest version.
i really like your framework; i prefer it over chatdev, autogen and crew-ai, but i have a question - in case i missed out on that and you covered this in a previous video: i your videos, you always use open ai, but i am using ollama to host my own llm's with lite llm as a reliable open ai api endpoint wrapper; can i run your agent swarm framework with my own model hosted in ollama and maybe lite llm as the open ai api endpoint?
@@vrsen thank you for the quick response! so i think i did not miss that out on a previous video, after all... thank you for clarifying; still like your framework, it seems very sophisticated!
Great work for one person project! Can you made a comparison video Crew AI vs Autogen vs Agency Swarm ? It might be very intereting to compare you personal project with this two with focus on design and possible extensibility.
Actually comparing to CrewAI is simple. As they are not true ai framework. They require to describe by code tasks , join tasks to agents, join tools to tasks)) I wonder where AI is here? If they are a true AI framework, then why can't they decide this without me) AgSwarm does this automatically based on agent's descriptions. Right?
the same with tasks. I have to describe tasks by code? If they are a true AI, then why can't their manager decompose my task by himself? ) Fmpv CrewAI is a pseudo AI framework. True framework allowed to operate only by system prompts, descriptions, tools and other declarative information. The rest of work AI should do automatically
Your thinking and implementation is just remarkable. I have idea in mind. Maybe to train the genesis for a while and build a library of agents and tools. and when user asks for something it will download the necessary agents and tools to make the agency creation faster and more accurate.
Good work, I'm trying but doesn't work for me (Ubuntu 22.04 and Windows 11), I'll try again with docker. It gets stuck when Developer agent try to execute create-next-app.
I played around with your Agency Swarm a bit, but the Instruction sets seem longer and more complicated than they should be. I think CrewAI has a slightly clearer method of defining the sequential tasks, so you know what it will do, whereas your method allows for more open-ended tasks. Honestly, though, I couldn't really get it to do anything "serious", and the costs were quite high (gpt4), considering that lots of the "internal communications" could probably be replaced with hard-coding for specific tasks.
You can hardcode as much logic as you need inside your agent's tools. I guess Crew AI is easier to get started with, but I would never recommend it in production. It does not have any type checking, nor error correction. (Watch my last short) "The process" is just a hardcoded prompt template within their system that takes away control from you. You can easily copy and paste it into agency swarm.
@@vrsen If you still see these comments - creating the Tools is something that isn't for "beginners". I tried, and failed (I'm not a pure SWE, so I suck). However, I think a basic tool would be a pseudo-memory function of simply saving things to TXT files. Like, have the agent save their results to a (temp even?) TXT file that can then be read/accessed by another agent asynchronously. It should be easy to implement, and is a pseudo RAG/Buffer/Memory trick that would make them super-useful! The other Agents would then need to have Read access (easy). Perhaps have a Summarizer agent always ensure that there is a short Summary of what's in the TEXT first (so as not to overload the context).
ПРивет. Хотел бы обратиться к тебе за помощью в разработке решения на базе которого можно интегрировать GPTs (именно свой обученный плагин) на свой сай в качестве чата, без каких либо сторонних решений.
Hey, I had an idea! What if we added a feature where users could send in an image of the layout they want, and then the AI creates everything for them? We could even take inspiration from layouts on Behance or Pinterest and give it a try! What do you think?
Great idea! I think I will try to create this once the vision model is available in assistants api. Right now it’s only possible to use it through a tool, which complicates things a bit
@@vrsen I was wondering if we could explore using Google Gemini. They offer a free API with vision capabilities, so maybe it could be a helpful tool for us!
I will answer as many comments as I can tomorrow. For now, stay tuned in our discord! ⚡(link in description) ☝
You're actually creating a whole new industry. Amazing, next-level stuff as always.
Thanks Paul!
Future developers are not just programmers; they are rather professional AI engineers who know how to use AI for their projects.
wow. After playing with CrewAi, Langchain and Autogen... this blew my mind.
Amazing work man. Please continue adding functionalities, you're on to something massive.
Best content. Real, no fluff! Thank you so much for not just making the videos but sharing your work!
You are changing the world. My world for sure.
Much love!
Okay why not add another agent that takes screenshots and uses gpt 4 vision with some saved images of great landing page designs to communicate with the developer to to create a much better design
Awesome. Would be really helpful if you could show a step by step for how to get set up from the very beginning. I got stuck on installation on Windows and can't get anything up and running
Your content is great and so straightforward! Thanks for these videos! 🙌
Thank you so much for this video. Super helpful and well-presented. Your time and effort is most appreciated!
Can it work with local LLM?
Awesome work Arsen and VRSEN team
If each agent has their own tools folder, how would you handle common tools? Through a fascade?
You can import them from the other agent's folder using standard python imports. However, if there is a need for this, most likely you are not splitting your agent roles right.
I suspect that you'll spend more time fixing and re-trying than doing it from scratch yourself... Not the right time for this stuff.
You don't understand what about is this game changing stuff. It's not about how quickly you create website like this, it's about finding new method. In few years it will be autonomous agency. But someone has to start developing. So your comment only show your old mindset. You live in old world.
Very cool…have you thought about somehow getting it to run inside a container? While I’m technical there’s no way I can adjust the code files…as you have…as not a pure dev….so won’t be testing this unless it’s on a dedicated machine…
can we use it with other llm's? like gemini or mixtral?
No, this framework is made intentionally on top of the latest OpenAI assistants api. Please refer to FAQ in our discord for details and discussion
@@vrsen This is a huge flaw to Agency swarms. I use Crew AI with Groq which delivers near instant responses with the speed being at real time I am able to to have my agents to talk to me in real sounding human like voices using text to speech platforms. Yet to experience a single hallucination.
Can it use your own templates that you have created and just build on those and if so how do I do that?
I love this. Could you maybe do a video on how to do it with local llms like lm studio?
Man your just blowing out my mind this is amazing now i think i need to start leveraging the skills in python
Awesome upgrades! 👍
Thanks Mike!
What program do you use to make the codes? Was it VS Code?
Wow! Feeling so proud! 😮
how much did it cost in api costs to run the ai agent swarm and build the website shown in this video? In next video, can you also please show how to use gpt 3.5 to save costs on api?
I will check next time, thanks for a suggestion
In your opinion, are openai assistants better than langchain's custom agents? The impression I have is that the assistants already had all the tools ready while the agents would have to choose and manage vector database, retrieval and memory resources.
Yes, there are many reasons why assistants are better. The biggest one is function calling. It makes your agents a lot more reliable in production. Creating tools from scratch is not a big issue, especially with genesis agency. It also gives you a lot more control
How about a system to streamline lead generation and qualification, aiming to replace the traditional Sales Development Representatives (SDRs) role? Researching leads can be time-consuming, so automating this process could provide a solution by delivering pre-qualified leads ready for outreach.
I have no idea why, but I haven't been able to get your code working, I keep trying as it's massively cool. Just trying to run the `agency-swarm genesis` command, I get:
usage: agency-swarm [-h] {create-agent-template} ...
agency-swarm: error: argument create_template: invalid choice: 'genesis' (choose from 'create-agent-template')
That's after having pip uninstall agency-swarm, then pip install agency-swarm, to ensure I'm running the latest version.
same issue
Really impressive, have you considered utilizing something like codium for unit testing? there approach seems to improve outputs.
Can you do it for open source llm's ??
i really like your framework; i prefer it over chatdev, autogen and crew-ai, but i have a question - in case i missed out on that and you covered this in a previous video:
i your videos, you always use open ai, but i am using ollama to host my own llm's with lite llm as a reliable open ai api endpoint wrapper; can i run your agent swarm framework with my own model hosted in ollama and maybe lite llm as the open ai api endpoint?
No, this framework is made intentionally on top of the latest OpenAI assistants api. Please refer to FAQ in our discord for details and discussion
@@vrsen thank you for the quick response! so i think i did not miss that out on a previous video, after all...
thank you for clarifying; still like your framework, it seems very sophisticated!
Hi, is that Visual studio code that you are using?
I am wondering too
Doenst Sentra Ai do the same thing? How is this in comparison to that platform?
This looks way more complicated.
Great work for one person project! Can you made a comparison video Crew AI vs Autogen vs Agency Swarm ? It might be very intereting to compare you personal project with this two with focus on design and possible extensibility.
Actually comparing to CrewAI is simple. As they are not true ai framework. They require to describe by code tasks , join tasks to agents, join tools to tasks)) I wonder where AI is here? If they are a true AI framework, then why can't they decide this without me) AgSwarm does this automatically based on agent's descriptions. Right?
the same with tasks. I have to describe tasks by code? If they are a true AI, then why can't their manager decompose my task by himself? ) Fmpv CrewAI is a pseudo AI framework. True framework allowed to operate only by system prompts, descriptions, tools and other declarative information. The rest of work AI should do automatically
Your thinking and implementation is just remarkable. I have idea in mind. Maybe to train the genesis for a while and build a library of agents and tools. and when user asks for something it will download the necessary agents and tools to make the agency creation faster and more accurate.
Good work, I'm trying but doesn't work for me (Ubuntu 22.04 and Windows 11), I'll try again with docker. It gets stuck when Developer agent try to execute create-next-app.
Awesome! Is it possible to use this with Azure OpenAI service?
Not yet, but I will add it later
@@vrsen , sounds great! Would love to see video about how to make it work when it's added!
Thank you!
crazy how this video only has 13k views this is by far the best agent framework and I've seen and tried them all - super smart and user friendly setup
I played around with your Agency Swarm a bit, but the Instruction sets seem longer and more complicated than they should be. I think CrewAI has a slightly clearer method of defining the sequential tasks, so you know what it will do, whereas your method allows for more open-ended tasks. Honestly, though, I couldn't really get it to do anything "serious", and the costs were quite high (gpt4), considering that lots of the "internal communications" could probably be replaced with hard-coding for specific tasks.
You can hardcode as much logic as you need inside your agent's tools.
I guess Crew AI is easier to get started with, but I would never recommend it in production. It does not have any type checking, nor error correction. (Watch my last short) "The process" is just a hardcoded prompt template within their system that takes away control from you. You can easily copy and paste it into agency swarm.
@@vrsen If you still see these comments - creating the Tools is something that isn't for "beginners". I tried, and failed (I'm not a pure SWE, so I suck).
However, I think a basic tool would be a pseudo-memory function of simply saving things to TXT files. Like, have the agent save their results to a (temp even?) TXT file that can then be read/accessed by another agent asynchronously. It should be easy to implement, and is a pseudo RAG/Buffer/Memory trick that would make them super-useful! The other Agents would then need to have Read access (easy). Perhaps have a Summarizer agent always ensure that there is a short Summary of what's in the TEXT first (so as not to overload the context).
@@zeevabrams I am doing this using Bubble
Awesome video as always VRSEN! How much did making this video cost in OpenAI expenses?
I'll track this in the next video
This is awesome 😍
How much did this whole process approximately cost OpenAI-API-wise?
I hope I find the time to dive into your awesome framework 😊
can i use it with local running llm?
No you can't
Great video! How much did this webpage cost you in API costs?
I'll check in the next video
Vrsen the GOAT:)
I think you do most of the work, and ai does some of the work⚡
Not clear for me this video, genesis? what's that?
Master, can we create a Ai agents that creates Ai agents?
Awesome👍👍👍!
The agency swarm legend is back! I always click so quickly on your videos 🙂
Thanks man! I’m trying to post a lot more, hence the new format
That's great man! I'm loving the new format! Keep up the great work
what was the old format?@@Skhulile84
Awesome:)
dayum son,where did you find theeesee (music)
Go boss!!!
Phenomenal
ПРивет. Хотел бы обратиться к тебе за помощью в разработке решения на базе которого можно интегрировать GPTs (именно свой обученный плагин) на свой сай в качестве чата, без каких либо сторонних решений.
Hi, feel free to reach out at www.vrsen.ai/ !
I be waiting for your videos to drop like a Jordan sneaker release😅😂
Likewise! 😀😀😀
Thank you, guys, I really appreciate your support. ♥ I'm trying to post more often, hence the new format
Hey, I had an idea! What if we added a feature where users could send in an image of the layout they want, and then the AI creates everything for them? We could even take inspiration from layouts on Behance or Pinterest and give it a try! What do you think?
Great idea! I think I will try to create this once the vision model is available in assistants api. Right now it’s only possible to use it through a tool, which complicates things a bit
@@vrsen I was wondering if we could explore using Google Gemini. They offer a free API with vision capabilities, so maybe it could be a helpful tool for us!
I already have this system set up using crew AI with Claude 3
Please, zoom in, what you are typing is very small. It is hard to see and read.
Got it, thanks
awesome stuff as always - but please please lose the crappy background music. so distracting
Wasted 15 minutes