Today is quite interesting. Two well - known companies each published an article, arguing about whether to use a multi - agent system.

Today is quite interesting. Two well - known companies each published an article arguing about whether to use a multi - agent system.

Claude's official Anthropic: How to Build a Multi - agent System
Devin's official Cognition: Don't Build a Multi - agent System

The core point of contention is: Should the Context be shared or separated?

Claude's view is that the essence of information search is compression. The context of a single agent is limited. When facing infinite information, if the compression ratio is too large, there will be distortion.
It's like no matter how capable a boss is, it's impossible to handle everything. One still needs to hire people to solve problems.
Through a multi - agent system, the boss asks different agents to study and report on key points respectively, and finally integrates them. Since each agent has its own expertise and diversity, it reduces the single - path dependence phenomenon. In practical terms, the multi - agent system also outperforms a single agent by 90%.

This is the victory achieved through collective wisdom and collaboration.

Devin's view is that the inconsistent contexts of multiple agents can lead to information fragmentation and misunderstandings. The information they report to the boss is often full of contradictions.
Moreover, many times, each action of an agent depends on the result of the previous step. However, multiple agents usually communicate with the boss separately and lack communication with each other, which easily leads to contradictory results.
This reflects the integrity and efficiency of individual wisdom.

Looking at the views of both sides, whether to use a multi - agent architecture is very much like the choice of running a company in human society.

A one - person company or a multi - person company?

In a one - person company, a person's mental capacity, physical strength, and time are very limited.
The advantage is that the communication cost of a one - person company is zero, and all time can be used efficiently.

In a multi - person company, the more people there are, the higher the communication cost, the greater the management difficulty, and the lower the overall efficiency.
But because there are more people, more mental capacity, and more physical strength, there is a possibility of more overall value output.
Designing a multi - agent system is very difficult, which is actually normal, just like running a company. It's hard.
The difficulty lies in establishing an effective collaborative system.
Moreover, the collaborative systems required for 1 person, 3 people, 10 people, 100 people, and 1000 people are quite different.

Referring to human history, relying on collective wisdom, humanity has achieved exponential development of civilization in modern times.
The collective wisdom of multi - agents may be the sprout for AI to achieve exponential development after the Scaling Law gradually slows down.

Regarding context, human collaboration has not been able to achieve perfect context management so far.

This reminds me that software engineering has never been about pursuing perfection, but about continuous iteration.

via AI Exploration Station - Jike Circle (author: orange.ai)