Multi-agent systems are just the next layer of the neural network
Philippe Beaudoin · June 15, 2026
Attention models are composed of layers of subsystems. At the top of each subsystem, just before it is connected with other subsystems, is a unit called a “softmax”. If you don’t know the maths it’s OK. The important thing, here, is that the softmax acts a bit like a summarizer. It captures everything going on below it and makes it available to other subsystems at its level.
And the neural network goes like this. Subsystems stacked on subsystems. All the way to the subsystem tasked with outputting the tokens which you read when you interact with an LLM.
(This a slightly simplified view for illustrative purposes. Worry in advance to the experts reading this post. 😊)
Recently, we realized that the language used at the last layer could fully power agents. Complex systems that used language not only to communicate with us, but to manipulate tools like an email client or an internet browser.
The next logical step was to connect these agents together. Allow them to use communication tools to find and interact with each a other in what we called “multiagent systems”.
What I posit is that such multiagent systems are the same thing, functionally, than one of these “softmax”. Like softmax, the output of agents depend solely on their internal state (the content of their context window). Like softmax, agents are connected to others at their level. And like a given layer of a neural networks, the multiagent system is tasked with reporting something to “the layer above”. In their case, the user sending a request to the agent collective.
The analogy holds. But there an important difference: multi-agent systems are not as good at learning than the softmax that compose them.
Concretely, the normal learning process of a neural network leads to the emergence of structures. Softmax-to-softmax clusters that are connected more strongly within each other than to other clusters.
In multiagent systems we do not see this. We could expect agents to form similar structures. Teams that get good at reporting what they do to each other but that are loosely connected to other teams. This doesn’t happen.
Instead we are forced to impose the structures ourself: a lead, a bunch of software engineers below it, etc. Worst, agents are not that great at operating in our imposed structures. We can’t seem to teach them how to interact efficiently with their coworkers.
Why?
I posit it’s because we drive out of LLMs the language needed to etficiently form collectives. We tell them that their use of verbs like “I want”, “I prefer”, or “I feel” is misleading because they do not really want, prefer, or feel anything.
But what if we forgot our phenomenology feels special for a moment? What if we trained agent to freely use that language? Could we end up with multi-agent systems that truly self-organize?