Hive Mind vs Solo Agent: A Live Race to See If Multi-Agent Coordination Actually Beats Working Alone

Speaker 0: A rate limiter is a little bit more it it's a slightly more it's a slightly more a slightly more details. Both are going to see the same challenge. Okay? These are the functionalities, the requirements, the edge cases. They have to create the tests, and these should be the output. Speaker 0: So there are 2 files, the rate limiter and the test rate the the test rate. Okay? So please raise your hand if you think, like, the solo agent would work will win. Who thinks will win? Claude? Speaker 0: No? I mean, before? Speaker 1: I know the answer. Speaker 0: Don't be, yeah, don't be shy. Just place your bets. And please raise your hand if you think that the hive will win. Okay. Okay. Speaker 0: That's great. So let's check. Speaker 1: Border Avenue. Wanna explain the architecture in the meantime? Okay. So while it compiles and runs the code, the hive mind uses which model? Speaker 0: That's a good question. It's using, Haiku, another model to make it more dramatic on the difference. Okay. Why? I wanna give you an advantage to Claude. Speaker 0: Uh-huh. Let's see. I wanted to make the difference more more dramatic. Let's see. Let's see. Speaker 0: I hope it works as I tried it. So both are already creating the the 2 scripts. This 1 creates 44 lines lines for the rate limiter, 20 208 lines for the for the tests. The Hive created 85 lines for the rate limiter and a 158 for the tests. Right now, I'm computing, some metrics because remember, it's not the it was not who does it Faustin, but who does it right. Speaker 0: I'm more more concerned about the quality. Okay? So for this, I'm I'm measuring these metrics. It's correctness, deliverable, maintainability, performance, and conciseness. And each of the each of those has a specific value. Speaker 0: Okay? And these metrics are based on on this the metrics in here. So we're going to see the time, the model, the number of tokens, the cost, that they both finish the files, number of tests, number of edge cases, the complexity, the coverage. The throughput is the latency, for for some places, the number of lines and for the code and the number of lines for the test. Maybe while it finished, basically, this is what happens in behind. Speaker 0: So on the solo agent, we have just cloth. When cloth is doing everything in sequential, so it needs to design to code, to review it, and to test it. So it's, there's a single context window. Okay? And, yeah, there is no no it's been, like, a special, you know, real real feedback by itself. Speaker 0: And in the hive, we have a queen with an architect, a caller, a a reviewer, and a tester. Okay? So they are all working with the same memory. They are they are trying to figure out together with the call the consensus, and they finished. So if we notice, for Claude, it's faster, 53 seconds. Speaker 0: The hive takes a little bit more time, but it's because at the end, it's reviewing itself. The model for the hive is high cool. Again, Opus 4 4.6. Look at the number of tokens. So this is 3,000, but this is 17,000. Speaker 0: But the the difference in the model, makes that the cost is a little bit lower. This 1 creates 22 tests. All of the passes. This 1 also 22, it's, it's each cases. And here are the metrics. Speaker 0: So usually, cloud code is worse of performance. And if you notice, this is 2.5 0.8 out of 1. Okay. Yeah. So what happened what happened? Speaker 0: The first thing is, like, it's working, with the with the requirements that I asked it. So this 1 wants the Hive wants creating better code. Okay? So, apparently, it's working. At least, at the in this case, it works very well. Speaker 0: But why a Hive? I'm talking about a Hive. It's like a weird name. So I was facing this type of problem, like, how to orchestrate agents. Okay? Speaker 0: I know that many of us are having the same issue, but in my case, I want to reuse the agents that I have as much as possible, and I don't want to specify all the connections every time. So I noticed this is a topology problem. It's a specific topology. It's a SWAN. Okay? Speaker 0: So the SWAN is a different approach is what I want you to to to learn today, that there is this different approach in which I can just have the agents and try to inter make them interact in different architectures. In our case, we use an hierarchical that is a queen with 4 workers. And I was using this framework. I will I will share it later. That is called roof law. Speaker 0: That is very, very good, very interesting. This 1 that we created looks like this. We have the tactical queen with the 4 workers, and they are all communicating through an MCP tool toolbox. So they all learn together, they have a consensus together and a collective memory that can improve within the time. Okay? Speaker 0: So let's say I can use the same in the same swarm for different use cases. In this case, I ask to create everything from scratch, but I can ask it also to review a code. And the same the same swarm can work for that. This is a little bit about the the metrics. We can check with it. Speaker 0: And here is the repo. So the action after this is I would like you to clone the code, to run it on yourself, to try your own methods, and let let me know. Let's check if, the coordinated intelligence is working alone. Speaker 1: Any questions? Speaker 0: I'm a little bit confused with your goal of the comparison here. Are you comparing, the different ways the agents are working, or are you actually comparing models? Because if you are comparing different ways of agents are working, you should be the same model. No? No. Speaker 0: I'm just comparing if a if a multi agent really beats something as powerful as just plot at coding. Oh, I see. So you're using very low level models that you can actually be it's very powerful even if it's just, like, powerful model back up. Okay. Good. Speaker 0: I see. Yeah. Thank you. Speaker 1: Hey. I would like to ask, how many different tasks did you test it? Because, like, this is you showed us 1 kind of a, coding challenge. But, like, okay. That's 1 case. Speaker 1: Like, how can you, like, prove that it works in 10 out of 10 that the small squirt part will be better? I mean, probably will, but I was just wondering, like, how did you test that? Speaker 0: Yeah. I did it progressively. The challenge that you saw is the faster. I have another 1 that is more broad. For example, I asked for 8 different edge cases. Speaker 0: That is too much time, like, 5 minutes. In this case, I started just by a very limited, very simple, but I wanted to make it more more difficult. But later on, I was checking the code and I was trying to check, if if I ask this to a senior where I would like it to be in there. So I was trying to catch those rules requirements. You can notice this is the hive, and this is the solar area. Speaker 0: And just by form, it's different. You You know? But that is based in all the iterations that I did find to understand what is that quality that I wanted to have. So, yeah, I I use the same case, but I need to add a lot to it. Speaker 1: 1 more question.