I enjoyed the step by step simple example. It’s a great overview for beginners.
I was thinking about it though and wanted your opinion. If you were going to build this particular scenario for real you’d have at least 2 agents collaborating would you not.
The web tool would likely be handled by a “Research Agent” while the other 2 tools probably live in a single “Personal Assistant” agent.
I know you were trying to keep this a simple, “this is why and where you’d use agents” example but I thought I’d add a bit of more context here in the comments.
Of course please disagree if you feel that all 3 tools would indeed like in a single Agent - in this case a “Personal Assistant” agent.
Good point, it can indeed make sense to create multiple agents.
I don't really have a strong opinion in this case, my intuition would be that it shouldn't make much of a difference, as the tasks don't require parallelization and/or aren't compute heavy. But I would definitely try out both solutions and compare them.
What's your personal experience with single- vs. multi-agent systems so far?
Well for full disclosure I’m building a multi agent system so I’m coming from this with a definite point of view.
I’d always separate them because if you start piling tools into one agent where does it stop?
This is a super simple example but if you were to build it this way in real life you’d lose all flexibility. You’d have to edit or create a new agent for even the smallest change. How about choosing web search tools to use as an example. To do that in this design gets messy really fast but if the web search tools are contained in a “Research Agent” (pick your name) it would be a simple change the prompt or a configuration depending on how your system is designed.
I think we need to think of agents (and I know this is controversial) more as analogous to human workers. Though we’d have computational agents and human stand in agents (amongst a few others that aren’t relevant to this example) and each would have atomic units of responsibility.
It does make the creation of a simple task theoretically (ok definitely) more difficult up front but the gains far outweigh the overhead imo since systems never stay simple.
I guess it really depends on the scope of the project and what's planned for the future. For tutorials and such, I usually like to keep it simple, as most people haven't even built a single-agent system yet. But if it's living project that is expected to grow in complexity, then I completely agree that we should think about it in a more modular way.
Agreed. I was hoping to add some let’s call it production world context as an adjunct to what I feel is an excellent beginner tutorial on agents. I think if I were to have 1 critique of the post it would be to add a bit at the end talking about this type of thing again assuming we’re in basic agreement that if we’re building real world systems we need to engineer them more like adaptive systems. I mean if not what’s the point of agents any way? ;)
However don’t take this as an actual criticism of the post. I thought it was an excellent beginner overview. One of the better I’ve read. I see questions that this would answer on the Reddit Agent groups multiple times a day so even though it feels like we’re way beyond this the reality is most people still don’t even really know why to use an agent over a basic set of function calls.
PS. By creating atomic agents as our capabilities grow we can apply more agent specific tuning. Example: learn that for recipe search the xyz search engine is best vs the abc engine for research paper search. I know I’m pushing the boundaries given the example in the post but just wanted to add a bit more why atomic agents makes sense from the get go and this is only 1 scenario.
Super excited to see the article live!! I was glad to work with you on this one Pascal :)
I enjoyed the step by step simple example. It’s a great overview for beginners.
I was thinking about it though and wanted your opinion. If you were going to build this particular scenario for real you’d have at least 2 agents collaborating would you not.
The web tool would likely be handled by a “Research Agent” while the other 2 tools probably live in a single “Personal Assistant” agent.
I know you were trying to keep this a simple, “this is why and where you’d use agents” example but I thought I’d add a bit of more context here in the comments.
Of course please disagree if you feel that all 3 tools would indeed like in a single Agent - in this case a “Personal Assistant” agent.
Cheers,
Christopher
Good point, it can indeed make sense to create multiple agents.
I don't really have a strong opinion in this case, my intuition would be that it shouldn't make much of a difference, as the tasks don't require parallelization and/or aren't compute heavy. But I would definitely try out both solutions and compare them.
What's your personal experience with single- vs. multi-agent systems so far?
Well for full disclosure I’m building a multi agent system so I’m coming from this with a definite point of view.
I’d always separate them because if you start piling tools into one agent where does it stop?
This is a super simple example but if you were to build it this way in real life you’d lose all flexibility. You’d have to edit or create a new agent for even the smallest change. How about choosing web search tools to use as an example. To do that in this design gets messy really fast but if the web search tools are contained in a “Research Agent” (pick your name) it would be a simple change the prompt or a configuration depending on how your system is designed.
I think we need to think of agents (and I know this is controversial) more as analogous to human workers. Though we’d have computational agents and human stand in agents (amongst a few others that aren’t relevant to this example) and each would have atomic units of responsibility.
It does make the creation of a simple task theoretically (ok definitely) more difficult up front but the gains far outweigh the overhead imo since systems never stay simple.
Thoughts?
Christopher
I guess it really depends on the scope of the project and what's planned for the future. For tutorials and such, I usually like to keep it simple, as most people haven't even built a single-agent system yet. But if it's living project that is expected to grow in complexity, then I completely agree that we should think about it in a more modular way.
Agreed. I was hoping to add some let’s call it production world context as an adjunct to what I feel is an excellent beginner tutorial on agents. I think if I were to have 1 critique of the post it would be to add a bit at the end talking about this type of thing again assuming we’re in basic agreement that if we’re building real world systems we need to engineer them more like adaptive systems. I mean if not what’s the point of agents any way? ;)
However don’t take this as an actual criticism of the post. I thought it was an excellent beginner overview. One of the better I’ve read. I see questions that this would answer on the Reddit Agent groups multiple times a day so even though it feels like we’re way beyond this the reality is most people still don’t even really know why to use an agent over a basic set of function calls.
Cheers,
Christopher
No worries, I'm glad you provided additional context.
It would also be a great topic for a follow-up article.
Indeed. If you feel I can have any useful input let me know and keep doing what you’re doing.
Christopher
PS. By creating atomic agents as our capabilities grow we can apply more agent specific tuning. Example: learn that for recipe search the xyz search engine is best vs the abc engine for research paper search. I know I’m pushing the boundaries given the example in the post but just wanted to add a bit more why atomic agents makes sense from the get go and this is only 1 scenario.
:)