夜雨聆风AI Agent之间的互动方式：哪种协作方式更有效？
Diverse Interactions Between AI Agents: Which collaboration approach is most effective?
Slides:
https://github.com/createmomo/Notes-on-AI-Agents
之前的文章：
从 OpenClaw 看 AI Agent 的运作原理 理解Context Engineering
Previous Articles:
How AI Agents Work: The OpenClaw Case Context Engineering Explained
关于这篇笔记的说明：
主要参考了Hung-yi Lee老师2026年的机器学习课程（https://www.youtube.com/watch?v=mmPmNezjCi0），其中一小部分内容根据个人理解可能有微小的调整； 保留了最核心的内容，所以比较简短，但也尽量保留了知识的完整。所以适合快速的回顾复习，当然也适合新手快速理解。
About This Note:
This note is primarily based on Prof. Hung-yi Lee’s 2026 Machine Learning course (https://www.youtube.com/watch?v=mmPmNezjCi0). A small portion of the content may include slight adaptations based on personal understanding; Only the core concepts are retained, making this note relatively concise while still preserving the completeness of the knowledge. Therefore, it is well-suited for quick review and revision, and also for beginners to gain a fast understanding.
在面对一些任务的时候，可能一个模型已经不能很好的应对了。但是，与其训练一个更大更厉害的模型，可能还不如让三个模型一同合作来解决同一个问题。
When facing certain tasks, a single model may no longer be able to handle them very well. However, instead of training an even larger and more powerful model, it may be more effective to let three models work together to solve the same problem.
现在其实已经有很多与此相关的研究了，有的论文就表明，让多个Agent互相讨论，可能就会获得比单个Agent的时候更好的结果。
In fact, there has already been quite a lot of research on this topic. Some papers have shown that allowing multiple agents to discuss with one another may lead to better results than relying on a single agent alone.
但，讨论的方式也可以有很多种呀，那怎么样的讨论方式才会达到最好的效果呢？
However, there can also be many different ways for agents to discuss. So, what kind of discussion structure can achieve the best results?
在这篇论文中，模型与模型之间，Agent之间的互动是用有向图来表示的。其中，每一个node就代表一个模型或者一个agent。
In this paper, the interactions between models, or between agents, are represented as a directed graph. Each node represents a model or an agent.
除此之外，其实每一条边（edge）也是一个模型。
In addition, each edge is also a model.
比如，在这个图当中，针对一个任务，前面的node会先分别提出自己的方案；
For example, in this graph, when facing a task, the preceding nodes first propose their own solutions separately.
然后，edge上面的模型会提供自己的想法和建议；
Then, the model on the edge provides its own thoughts and suggestions.
最后这个node上的模型会综合前面的方案和建议，来提出自己的想法。
Finally, the model on the last node synthesises the previous solutions and suggestions, and then puts forward its own idea.
所以我们可以看出，最后这个节点的工作并不是把前面的内容做简单的拼接，而是也会有自己的思考在里面。
Therefore, we can see that the role of the final node is not simply to concatenate the previous content. Rather, it also involves its own reasoning and judgement.
那，在了解了基本概念之后，我们就可以看一下这篇论文到底尝试了哪些不同的讨论方式。
Now that we have understood the basic concept, we can look at the different discussion structures explored in this paper.
比如，这种相对比较简单的一个传一个的方式，也就是接龙的方式。
For example, there is this relatively simple one-after-another structure, which is similar to a chain.
还有的方式是这种树状结构的方式。
Another approach is this tree-like structure.
不过这里需要注意的是，在这里是有一个node节点产生想法，然后再传给其他人去发散想法，从而提出多个方案。
However, it is worth noting that, in this case, one node first generates an idea, and then passes it on to others so that they can expand on it and propose multiple solutions.
这样的方式其实有一点反直觉，因为我们的直觉可能是由一棵树的最底层的节点产生想法，然后层层往上汇报，直到汇报到最上面的节点。但在这里不是这种情况。
This approach is actually a little counter-intuitive, because we might instinctively imagine that the nodes at the bottom of a tree generate ideas first, then report upwards layer by layer, until the information reaches the top node. But that is not the case here.
总之，发散出来多个方案后，最终会有一个隐藏的node（没有在图中显示），会把所有的方案总结汇总起来给出一个最终的方案。
In short, after multiple solutions have been generated, there will eventually be a hidden node, which is not shown in the figure, that summarises and integrates all the solutions to produce a final answer.
那除了CHAIN和TREE的结构，还可以有更复杂的结构。比如这种Mesh Graph，这种图的特点就是所有节点两两之间都是会有关联性的。比如我们可以看出，会有1个节点会把想法传给4个节点，然后会有1个节点传给3个节点等。反正两两节点之间是会有edge相连接的。
Apart from CHAIN and TREE structures, there can also be more complex structures. One example is this Mesh Graph. The key feature of this type of graph is that every pair of nodes is connected in some way. For instance, we can see that one node passes its idea to four other nodes, whilst another node passes its idea to three nodes, and so on. In short, there is an edge connecting each pair of nodes.
如果我们再往后看呢，会看到类似于神经网络的样子。虽然看起来像神经网络，但实际上每一个节点都是一个单独的神经网络模型（毕竟现在的模型都是用神经网络做的嘛）。
If we look further, we can see a structure that resembles a neural network. Although it looks like a neural network, each node is in fact a separate neural network model, since today’s models are, after all, built using neural networks.
那我们再往后看，可以看到，随机生成的Graph。这种图实际上就是从前面的Mesh图随机做pruning而得到的。
Looking further again, we can see a randomly generated graph. This type of graph is actually obtained by randomly pruning the previous Mesh graph.
那这些不同的讨论方式，都有怎么样的效果呢？
What are the effects of these different interaction structures?
在这个图中，横轴大概是指，我们现在动用了多少“人力”，也就是动用了多少Agent来完成任务；竖轴表示的是完成任务的质量分数。这里的分数是在4种不同的任务上，最终取平均值所获得的分数。
In this figure, the horizontal axis roughly represents the amount of computational “manpower” being used, namely the number of agents involved in completing the task. The vertical axis represents the quality score of the final output. This score is calculated as the average performance across four different tasks.
从左上的图，我们可以看出，效果最不好的，是那种一个传一个（Chain）的方式。
From the figure in the upper-left corner, we can see that the least effective approach is the sequential one-agent-after-another structure, namely the Chain structure.
那其他方式呢？论文中发现了，其实不同的任务，最适合的讨论方式也是不一样的。那到底什么方式最适合什么任务，就只能是case by case了，这里面可能会有很多需要研究的地方。
What about the other approaches? According to the paper, the most suitable interaction structure may vary across tasks. In other words, which structure works best depends on the specific task under consideration. This also suggests that there is still considerable room for further research in this area.
那，从这个图中我们还可以看出，随着投入的agent越来越多，质量确实也会变得越来越好。但这种变好是有一个上限的。当对agent的投入超过某一个点之后，提升的效果就不会太明显了。
The figure also shows that, as more agents are introduced, the overall quality does improve. However, this improvement is not unlimited. Once the number of agents exceeds a certain threshold, the marginal gains become much less significant.
