I would say it hardly compares. This looks too basic for anything more than a demo. Perhaps in the future but for now I don’t get it. There are better state machine solutions out there
@@wesleymogaka swarm is simpler. Maybe for mvps. Langgraph is robust. Better for prototypes and production. Both can get the job done. Really down to preference and skill
So funny. This is how I program my custom agent state machines. I keep telling people agents are classes with LLM functionality (http/s requests). Nothing more... The desired end state/output is a game loop. Nothing more. I still code everything up my self and developed my own package. Works very well when switching between openai and olama
That was my thought today, why do we need to so disrupt this process? Why can't it be made simpler instead of more complex? I was going to start a little project, but I see someon did it already :D
I agree. Why not give a set of tools to the LLM and, through CoT or ReAct, allow it to decide the workflow? Almost all of the examples I see are one-sentence questions or queries. For example, here is a typical request/query I can give my demo: "Search the internet for PDF documents on garlic and websites that have information on garlic. Save all the information found to the Top drawer. Ingest all of the information found into the knowledge base.
It has been my experience that the LLMs can only handle so many tools before it becomes inconsistent and less reliable. The agent framework helps you break down those tool calls into smaller segments that are more reliable. It would be nice though if one agent could handle many tools so you didn’t need to make all these different “agents” to handle the tool calls.
LLMs get confused with too many rules, edge cases, etc. Having multiple agents help to focus that particular instance of the LLM to some specific task, allowing its context window to not be "contaminated" with tokens that can make the model confused. It also lowers overall complexity and helps with modularity since you can change a specific agent behavior without having to change the main one or even having to deal with unpredictable side effects of changing a big system prompt in which every token interacts with many others leading to unwanted changes in behavior.
in software development there are some important concepts to create a component: Single responsibility, Low coupling and testability. When you have a component with many responsibilities you cannot ensure it will work as expected. That is why software is preferred to have multiple single complements well tested and working than only one that has weird behaviors
swarm is counter-intuitive to AI. The AI is suppose to build, grow, refine, manage logical structures all on its own("behind the scenes") just like its learning and growing just by users inputs(prompts). Users aren't suppose to take this capability from AI, its suppose to do this in its "head" all while being prompted...probably very little productivity gains in actuality with swarm.
@@ceo.amai88 Absolutely. In programming, a "client" is specifically that part of the program requesting data from a server. It is the first hand in a "handshake". If someone with any programming experience sat down at that code, they would be baffled by the identification of an AI swarm as "the client". A good example of this is an MMORPG. The person downloading your "client" is the "client". That's bad enough. But imagine if you added an AI army "client" inside of the "client" for use by the "client". We are approaching a new world where "programming" is more about naming things in a way that keeps things organized rather than an expertise in the nuances of the target language.
@@ceo.amai88 The computer never cares what English is used to identify a variable. It all becomes 1s, 0s, and pointers for them. It is, however, bad tradecraft because another programmer (or even yourself after time has passed) must depend on an accurate description of what your variables are meant to do to be conveyed by their name. That's why calling any variable "X" is a cardinal sin among senior programmers. If he called that swarm "foo", that would be weird; if he calls it "client", he is imparting false information that can only make his life harder.
@@dirtydevotee hahaha, ok i get it. it's all about systemization/segmentation. and it could affect how you would scale your system and your (future) workflow. thanks a lot.
I lt would be too risky, I thought the very second when so many comments popped out after swarm news. questions why so many and not one or crew. Answer is simple, too much power in single unstable point lolz.
Thank you for line by line go through ! appreciated it !!!
Please create a video on comparison of several such frameworks, and their differences in approaches to solve the multi agentic problems.
How does it compare with langgraph, can we use it with another LLM?
Transferring control looks like edge connection from node a to node b in langgraph
I would say it hardly compares. This looks too basic for anything more than a demo. Perhaps in the future but for now I don’t get it. There are better state machine solutions out there
@@wesleymogaka swarm is simpler. Maybe for mvps. Langgraph is robust. Better for prototypes and production.
Both can get the job done. Really down to preference and skill
Thanks for sharing! Have you tried Reka AI models by the way? They just released a new 21B model.
I had similiar idea, in fact the concept is rather trivial. That's reasurring :)
reassurring in what way ?
The idea is on the surface. Everyone is doing the same thing :)
@@AlexK-xb4co me too, who's not
What application do you use for flowcharts?
excalidraw.com
I'm guessing Excalidraw
great run down
So funny. This is how I program my custom agent state machines. I keep telling people agents are classes with LLM functionality (http/s requests). Nothing more... The desired end state/output is a game loop. Nothing more. I still code everything up my self and developed my own package. Works very well when switching between openai and olama
i alternate between groq and llamafile(local) --- same idea
Brilliant. Can you teach or share code structure?
Has anyone heard if Semantic Kernel is going to be wrapping the swarm framework too?
Can you use this framework with other models like Gemini and open source?
It is literally two files module, you can write the things yourself; hope they add more features
How does this differ from something like CrewAI?
Thanks!
so...who's gonna be the first to create a swarm of o1-preview agents?
I guess o1 itself is a swarm of gpt-4o
What are the practical applications of this tech?
Automated system, integrated system or anything you can possibly imagine
thanks:)
That was my thought today, why do we need to so disrupt this process? Why can't it be made simpler instead of more complex? I was going to start a little project, but I see someon did it already :D
be careful, you might end up coding another tool in the line of langflow / langgraph / n8n / microsoft autogen / whatever else appeared recently
@@AlexK-xb4co "recently"? lol
AWESOME
I just don't get it. Why not give all the tools to a single agent instead if distributing it over several agents?
I agree. Why not give a set of tools to the LLM and, through CoT or ReAct, allow it to decide the workflow? Almost all of the examples I see are one-sentence questions or queries. For example, here is a typical request/query I can give my demo:
"Search the internet for PDF documents on garlic and websites that have information on garlic. Save all the information found to the Top drawer. Ingest all of the information found into the knowledge base.
It has been my experience that the LLMs can only handle so many tools before it becomes inconsistent and less reliable. The agent framework helps you break down those tool calls into smaller segments that are more reliable. It would be nice though if one agent could handle many tools so you didn’t need to make all these different “agents” to handle the tool calls.
LLMs get confused with too many rules, edge cases, etc. Having multiple agents help to focus that particular instance of the LLM to some specific task, allowing its context window to not be "contaminated" with tokens that can make the model confused.
It also lowers overall complexity and helps with modularity since you can change a specific agent behavior without having to change the main one or even having to deal with unpredictable side effects of changing a big system prompt in which every token interacts with many others leading to unwanted changes in behavior.
The rational is explained in the blog post: scalability and modularity
in software development there are some important concepts to create a component: Single responsibility, Low coupling and testability. When you have a component with many responsibilities you cannot ensure it will work as expected. That is why software is preferred to have multiple single complements well tested and working than only one that has weird behaviors
We move like a $Swarm, decentralized and leaderless
swarm is counter-intuitive to AI. The AI is suppose to build, grow, refine, manage logical structures all on its own("behind the scenes") just like its learning and growing just by users inputs(prompts). Users aren't suppose to take this capability from AI, its suppose to do this in its "head" all while being prompted...probably very little productivity gains in actuality with swarm.
This is like CrewAi or something similar
You called your Swarm variable "client". That's bad tradecraft. You should call it, "this_swam".
can you explain why? thanks
@@ceo.amai88 Absolutely. In programming, a "client" is specifically that part of the program requesting data from a server. It is the first hand in a "handshake". If someone with any programming experience sat down at that code, they would be baffled by the identification of an AI swarm as "the client".
A good example of this is an MMORPG. The person downloading your "client" is the "client". That's bad enough. But imagine if you added an AI army "client" inside of the "client" for use by the "client".
We are approaching a new world where "programming" is more about naming things in a way that keeps things organized rather than an expertise in the nuances of the target language.
@@dirtydevotee but why does it matter? does it matter to the programmer, or the computer trying to read/identify it?
@@ceo.amai88 The computer never cares what English is used to identify a variable. It all becomes 1s, 0s, and pointers for them. It is, however, bad tradecraft because another programmer (or even yourself after time has passed) must depend on an accurate description of what your variables are meant to do to be conveyed by their name.
That's why calling any variable "X" is a cardinal sin among senior programmers. If he called that swarm "foo", that would be weird; if he calls it "client", he is imparting false information that can only make his life harder.
@@dirtydevotee hahaha, ok i get it. it's all about systemization/segmentation. and it could affect how you would scale your system and your (future) workflow. thanks a lot.
is it me? or this seems like a super complicated if else statement?
not really seeing anything new here. Agent implementations in langgraph seem way ahead of this at the moment.
I heard its expensive lol
I lt would be too risky, I thought the very second when so many comments popped out after swarm news. questions why so many and not one or crew. Answer is simple, too much power in single unstable point lolz.
bro just yapping
People will do anything besides just learn graph
What’s graph
??
What's graph got to do with agents?
@@jonno081 what is graph even
LangGraph 💪🏾