If you have watched “Naruto”, you should remember Naruto Uzumaki’s combination technique: the Shadow Clone Technique combined with the Transformation Technique. One main body, and with a “bang”, a bunch of clones appear, each capable of playing different roles—some engage in direct combat, some gather intelligence, and some simply take hits to test things out, ultimately dissipating together, with all experiences returned to the main body.
In simple terms, Naruto’s combat power isn’t extraordinarily strong, but he relies on a complete set of “me and my clones forming a team” approach.
Today, we are going to discuss Multi-Agent Collaboration, which is somewhat like building a “real-world version of the shadow clone system” for AI.
However, we are in a worse situation: current large models do not yet have the technique of “experience sharing”, and their worldview is far from the level of the “final battle of Naruto”, so don’t expect a soul-piercing Naruto; we still need to start from the basics of “multiple clones working together”.
In the previous two articles, we have been using two extreme roles as analogies: one is Zhuge Liang, who has a world model and can weave fragmented information into a future prediction map; the other is Sheldon, who has only a day’s memory—high IQ, but after sleeping, everything from yesterday is wiped clean.
These two thought experiments correspond to two major shortcomings of current large models:
First, the world model is not stable enough. It does not have a stable, updatable, and self-consistent worldview like Zhuge Liang’s, which can accurately understand “what is happening now”, “how causality leads to the future”, and “how multi-party games will evolve”; it is more about interpolating correlations rather than genuinely extrapolating from causal structures.
Second, memory and self-updating are too weak. When the chat window closes, it is like Sheldon going to sleep—context is reset, and the model itself does not grow from chatting with you for three months.
Under this premise, if our strategy is only one sentence:
“Make the model a bit larger, a bit faster, and have a bit more knowledge.”
Then essentially, it is still about stacking computing power and data, rather than acknowledging structural shortcomings.
Thus, another more realistic path emerges: since we cannot temporarily create a perfect Zhuge Liang, let’s change our thinking—
Don’t expect one model to handle everything; break a large task into multiple roles and steps, allowing a group of “limited smart” agents to work together on something much larger than themselves.
This is very similar to Naruto’s approach: the main body may not be extraordinarily strong, but with enough clones, clear division of labor, and information feedback, it can achieve a “combined effect that exceeds the individual”.
Multi-agent systems essentially use “mechanism design” to compensate for the shortcomings of a single model in world modeling, memory, and self-supervision.
It is not about giving AI a cooler business card, but rather acknowledging: our current intelligence is still in the “organizational engineering” stage.
While multi-agent systems are considered cutting-edge, the most primitive version has existed since the abacus era.
Back then, without calculators, a ledger with thousands of entries relied entirely on people using abacuses. Humans can make mistakes, which is human nature. So what to do? They established a very simple yet elegant system: dual independent calculations.
The process is straightforward: Accountant A calculates the entire ledger from start to finish without consulting anyone else; Accountant B also calculates independently, without peeking. Finally, the two compare results: if they match, it passes; if not, they must recalculate, even involving a third person for verification.
On the surface, it seems like just adding another person. In essence, it employs the structure of “redundancy + independence + divergence triggering verification”:
Errors between the two are mostly independent; only if both happen to calculate the same wrong number can the error slip through; once there is a discrepancy, it automatically triggers stricter checks, effectively adding a “Safety Agent”.
You can think of the dual independent calculations of the abacus era as the earliest “multi-agent consensus protocol”.
Today’s LLM council and multi-agent collaboration frameworks essentially bring this old-fashioned method into the AI world.
Let’s start with a very intuitive example: the open-source LLM Council project by Andrej Karpathy.
Its approach is: the same question is thrown to several different large models simultaneously—such as GPT, Claude, Gemini, Grok… Each model answers independently in a “closed room”, unaware of what others have written. Then, their answers are anonymously peer-reviewed and scored, with a “chair model” synthesizing opinions to produce the final version.
This is like a “multi-model review meeting”: each model is both an author and a reviewer, and ultimately, there is an editor to finalize it.
The benefit of this approach is:
Errors are no longer highly correlated; different models have different biases in data and alignment, significantly reducing the probability of making the same mistake; opinions are richer, with some models favoring frameworks and others focusing on details; the final “consensus result” is often more stable in high-uncertainty scenarios than any single model.
From an investment perspective, this essentially introduces the concepts of cross-checking and triangulation.
Of course, you can also use just one model and create a bunch of different agents internally, which brings us back to Naruto’s shadow clone approach:
Multi-agent division of labor and collaboration
Some projects have fully agentified their “investment research department”. For example, TradingAgents: Multi-Agents LLM Financial Trading Framework has built a virtual trading company:
-
One specializes in fundamental analysis
-
One monitors news and public sentiment
-
One studies technical indicators
-
One is responsible for comprehensive decision-making
-
One acts as a risk manager
Each agent does not need to “understand the whole picture like a god”; they just need to clarify the logic in their own area, and then through established processes and dialogue rules, assemble these judgments into an executable trading decision.
Another way to break it down is by “stance and style”, assuming from the start that they have differing opinions, even encouraging them to argue. This falls under multi-agent role-playing.
In investment scenarios, this is quite natural: you can define an extremely optimistic bull agent, focusing only on space, stories, and structural growth; an extremely conservative bear agent, only watching valuations, cash flow, and downside risks; a macro agent, looking at interest rates, exchange rates, policies, and liquidity; and a trading agent, mainly concerned with sentiment, positions, technical levels, and funding rhythm.
They can hold a “virtual investment committee” around the same stock or macro event, with a CIO agent making the final judgment. The value lies not in who wins the argument, but in forcing you to lay out the logic of different styles on the table, making it clear where the divergences are, rather than just seeing a vague “consensus opinion”.
Many decisions are difficult not because we lack a smart conclusion, but because we lack a clear “map of divergences”.
Taking it a step further, agents do not necessarily have to play the role of “employees”; they can also become “users” directly.
There is a paper dedicated to this, titled “LLM-Based Multi-Agent System for Simulating and Analyzing Marketing and Consumer Behavior”: it uses a group of LLM agents to simulate consumer decisions and social interactions under different marketing strategies.By treating hundreds or thousands of LLM agents as simulated users: each agent has different personalities, incomes, preferences, and social network positions. Then, they are given different advertising copy, discount strategies, and product changes to see how they react and how they influence each other in social networks.
While it is called “predicting user behavior”, it is more like creating a miniature society, throwing growth strategies and brand stories into it, and observing how it might evolve in this small world.
This cannot be precise to the percentage point, but it provides another very useful capability: before trying things in the real world, you have a relatively cheap “simulation battlefield”. Even if it only helps you eliminate some obviously disastrous options, it is worth it.
If the previous examples leaned more towards “rational decision-making”, there is also a line specifically addressing the “emotional intelligence shortfall”. Recently, a study titled MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems serves as a typical example. Its work can be simply summarized in one sentence:
It breaks down the task of “understanding others + reading the room + saying something human” into several specialized agents, andeven in a very small, few-turn dialogue scenario, it significantly improves dialogue quality through multi-agent collaboration.
In the design of MetaMind, there is a Theory-of-Mind agent that specializes in guessing the other party’s current intentions, emotions, and beliefs—such as whether “I’m fine” is genuinely fine or just a polite way of saying they are uncomfortable, and whether there is an implication; there is a Domain/Norm agent that filters these guesses through specific cultural and social norms: what is considered considerate and what is considered offensive in different contexts; and finally, there is a Response agent that generates a response that is both reasonable and socially acceptable based on the first two, and checks it for potential pitfalls before sending it out.
This means that even for a very short dialogue segment, MetaMind does not simply let the large model respond directly; instead, it allows several small agents to hold an internal “brainstorming session” before providing a response that is closer to human social sensibilities.
Experimental results indicate that this approach can significantly enhance the performance of large models in social reasoning and Theory-of-Mind tests, while also improving the subtlety and safety of responses in actual dialogues.
In other words:
Enhancing AI’s “emotional intelligence” does not necessarily require training a more mysterious and larger single model; it can also be achieved by breaking down social cognition into several specialized agents that collaborate in a small dialogue scenario.
At this point, we can wrap up.
If you remember the roles we mentioned at the beginning—Zhuge Liang with a world model, Sheldon with only a day’s memory, and Naruto who can use shadow clones—they actually represent three different technical demands:
Zhuge Liang represents a powerful, stable world model; Sheldon reminds you of the shortcomings in memory and self-updating; and Naruto tells you that when a single entity is not strong enough, “collaboration among many clones” can greatly amplify your limits.
At this stage, where single large models cannot yet achieve “understanding the world, having long-term memory, and continuous growth”, the line of multi-agent collaboration provides a very simple yet effective engineering answer: use redundancy, use independence, and use “divergence triggering verification” to minimize systemic errors, break down complex problems into smaller parts, and ensure that the overall intelligence level of the system is more stable than any single point.
From dual independent calculations in the abacus era to Karpathy’s LLM Council, to multi-agent investment research systems simulating a virtual research department, marketing teams creating a miniature market, and MetaMind breaking down social cognition. They may seem diverse, but the underlying logic is consistent.
If you view AI as an organization rather than a deity, then multi-agent collaboration is the best choice.