kinda like tinytroupe, except that the user can add to the conversation at any point in time. this could be useful for so many things, such as research agents using different models per their strength (math, logical reasoning, domain specific knowledge such as physics, chemistry, mechanical engineering, even law, ...) discussing a subject in the background, and the user can read everything in real time and in a history, and chat or add documents to add context.
Great video! Subbed right away! I’m running into problems with the window buffer memory when calling other workflows. I can’t seem to figure it out….sessionID is the error I get. I’m using a google sheet as a starting point. Any suggestions? Thanks.
@laurensveldman5180 sessionID is required for window buffer memory to work. You can set it to a unique value (e.g file id or user id). Alternatively session id can be passed from outside while invoking the n8n webhook and used in window buffer memory.
this is true, but there is also way more latency involved as well... double or triple, maybe even more latency... if there are functions or tools you use alot, they should be accessible by the main or router agent
@RamonTomzer True, latency is a key factor to consider, especially when dealing with frequently used tools. If the use case requires low latency a number of other optimizations are required.
good insight. this youtuber has an interesting take on the multi-agent setup: hierarchy vs sequential ... but I'm relatively new to all this, and assuming every use case is different, and will produce varying results ua-cam.com/video/YtUg8hesGTc/v-deo.html ... would like to hear our thoughts on this video
@SoloJetMan Yeah somewhat agree with the thoughts in the linked video. But it all boils down to the usecase. If you know exactly what order your agents need to follow and your demands won't grow swiftly in future, sequential can make more sense. Else the workflows can become really complex though with more control. But things can break there as well anyways. In multi agent, if we can maintain the good prompt and use the router agent just for the purpose of invoking correct agent tools, we should be good. It can easily accommodate future requirements and workflows are manageable. In any case, a single agent shouldn't be overloaded with a lot of tasks. 1 task = 1 agent. Another important thing to note is, the perf of multi agent also depends on the LLM being used. It gives good results with gpt4o. But not with llama (unfortunately). In future, this will improve and become more and more reliable, making multi agent even more preferred. At the end it's a tradeoff and depends on the usecase.
Creating a multi-agent workflow with Superwiser on top is not the best idea for real business solutions. It looks modern and effective on YT videos, but in reality it is highly unpredictable and works with errors.
@MariushAI Thanks for sharing your views. This is how I would summarize this: Sequential: Best for simpler, step-by-step tasks where reliability and predictability matter most. Multi-Agent: Best for complex, diverse tasks requiring parallel work and specialized expertise. Will create a separate video on when to use which with perf measurement.
@@FuturMinds The parallel work can be achieved by sequential multi-agent workflows with all complexity and specialized knowledge. With more control over the process and fully predictable results.
All of the YT content with the words "advanced", "secret", and "masterpiece" are cool and fun. However, not all workflows are "production ready".This is for testing purposes only. In real life and with real problems, you need full precision and full control.
Sure, thanks for sharing your thoughts. I believe none of the AI approaches are perfect and are evolving rapidly to become more reliable. Give that, a single approach is not suitable for all usecases. A number of things depend on the context. Will create a video soon on the indepth comparison of different approaches.
Fantastic video! I'm learning so much about n8n from your content. I'm really eager to collaborate with you.
Appreciate the kind words :)
Thank you, your content is top notch !!
@myscubajourney2895 Thanks, appreciate the kind words :)
very cool and great ideas, thought process, explanations !!!
Thank you! Cheers!
Meus parabéns pelo vídeo!
Thanks @fhelypg :)
again, very well done! would you be able to show how a "autonomous discussions between agents with user input" kind of scenario?
kinda like tinytroupe, except that the user can add to the conversation at any point in time. this could be useful for so many things, such as research agents using different models per their strength (math, logical reasoning, domain specific knowledge such as physics, chemistry, mechanical engineering, even law, ...) discussing a subject in the background, and the user can read everything in real time and in a history, and chat or add documents to add context.
@themax2go This sounds interesting. Let me think about this.
Great video! Subbed right away! I’m running into problems with the window buffer memory when calling other workflows. I can’t seem to figure it out….sessionID is the error I get. I’m using a google sheet as a starting point. Any suggestions? Thanks.
@laurensveldman5180 sessionID is required for window buffer memory to work. You can set it to a unique value (e.g file id or user id). Alternatively session id can be passed from outside while invoking the n8n webhook and used in window buffer memory.
this is true, but there is also way more latency involved as well... double or triple, maybe even more latency... if there are functions or tools you use alot, they should be accessible by the main or router agent
@RamonTomzer True, latency is a key factor to consider, especially when dealing with frequently used tools. If the use case requires low latency a number of other optimizations are required.
Great video
Glad you enjoyed it
good insight. this youtuber has an interesting take on the multi-agent setup: hierarchy vs sequential ... but I'm relatively new to all this, and assuming every use case is different, and will produce varying results ua-cam.com/video/YtUg8hesGTc/v-deo.html ... would like to hear our thoughts on this video
@SoloJetMan Yeah somewhat agree with the thoughts in the linked video. But it all boils down to the usecase. If you know exactly what order your agents need to follow and your demands won't grow swiftly in future, sequential can make more sense. Else the workflows can become really complex though with more control. But things can break there as well anyways.
In multi agent, if we can maintain the good prompt and use the router agent just for the purpose of invoking correct agent tools, we should be good. It can easily accommodate future requirements and workflows are manageable.
In any case, a single agent shouldn't be overloaded with a lot of tasks. 1 task = 1 agent.
Another important thing to note is, the perf of multi agent also depends on the LLM being used. It gives good results with gpt4o. But not with llama (unfortunately). In future, this will improve and become more and more reliable, making multi agent even more preferred.
At the end it's a tradeoff and depends on the usecase.
@@FuturMinds thanks for taking time to respond!
@SoloJetMan thanks for bringing this up. Will create a separate video on when to use which with perf measurements and comparisons soon.
Creating a multi-agent workflow with Superwiser on top is not the best idea for real business solutions. It looks modern and effective on YT videos, but in reality it is highly unpredictable and works with errors.
@MariushAI Thanks for sharing your views. This is how I would summarize this:
Sequential: Best for simpler, step-by-step tasks where reliability and predictability matter most.
Multi-Agent: Best for complex, diverse tasks requiring parallel work and specialized expertise.
Will create a separate video on when to use which with perf measurement.
@@FuturMinds The parallel work can be achieved by sequential multi-agent workflows with all complexity and specialized knowledge. With more control over the process and fully predictable results.
All of the YT content with the words "advanced", "secret", and "masterpiece" are cool and fun. However, not all workflows are "production ready".This is for testing purposes only. In real life and with real problems, you need full precision and full control.
Sure, thanks for sharing your thoughts. I believe none of the AI approaches are perfect and are evolving rapidly to become more reliable. Give that, a single approach is not suitable for all usecases. A number of things depend on the context. Will create a video soon on the indepth comparison of different approaches.
@@MariushAIat what point then should a MAS system be used
How it worth, can we talk about this system do you have email or phone?
@ParkersYourDad Feel free to send an email. You can find the email in the bio.