As argued in the previous chapter,
AI does not understand the world.
What it does is nothing more than smooth function approximation.
And yet,
we still want to use AI for decision-making.
At that point, we inevitably face the next question:
“Whose judgment should the AI make?”
There is no single answer to this question.
Decision-making is structurally fragmented
Traditional AI design has relied on an implicit assumption:
One intelligent model understands everything and produces the optimal answer.
This assumption is an illusion.
Because real-world decision-making is fragmented from the very beginning.
-
Sales judgments and legal judgments are different
-
Local optimization and executive optimization do not coincide
-
Short-term profit and long-term value often conflict
These are not cases of “missing information” or “noise.”
When roles and responsibilities differ,
it is natural that judgment criteria differ as well.
In other words, this is not an error to be corrected, but
a structural discontinuity that inherently exists.
Do not eliminate the discontinuity — preserve it
This is where many AI projects fail.
Different judgments are
-
averaged,
-
merged,
-
and flattened under the name of “optimization.”
What happens as a result?
A plausible-looking answer emerges —
but it belongs to no one.
Responsibility becomes unclear,
and no one can explain why that decision was made.
Multi-agent design takes the opposite approach.
It does not resolve discontinuities.
It embeds them directly into the system design.
That is the starting point.
Multi-agent systems are not “distributed understanding”
Here, one critical misunderstanding must be cleared up.
A multi-agent system is not
“multiple AIs collaborating to understand the world more deeply.”
No one understands the world in the first place.
What each agent possesses is not a worldview,
but a judgment axis.
-
An agent that only checks whether constraints are violated
-
An agent that maximizes efficiency or expected value
-
An agent that detects risk signals
-
An agent that stops the process and says, “This must be decided by a human”
They do not agree.
They are designed on the assumption that agreement will not be reached,
each raising objections from its own standpoint.
Consensus is not the goal
This is the most important point.
In multi-agent design,
Consensus is not a premise.
Consensus is merely a possible outcome.
Sometimes, stopping without agreement is the correct decision.
-
Constraints are violated
-
Risks cannot be explained
-
Decision rationales are not written into the contract
To prevent progress under such conditions,
agents are meant to disagree.
This is not failure.
The fact that judgments diverge is itself a sign of system health.
Abandon the illusion of collective intelligence
Multi-agent systems are not designed
to make the whole system “smarter.”
In fact, they do the opposite.
They intentionally abandon the idea of treating the whole as a single intelligence.
Judgments are decomposed.
Responsibilities are separated.
Conflicts are made explicit.
Instead of forcing AI to “understand” the world,
we directly project human decision structures into the system.
That is the essence of multi-agent design.
A small conclusion
AI does not understand.
Multi-agent systems do not create understanding either.
And yet, they matter because they preserve this:
Who disagreed, where, and why — as a structural record.
As long as that structure remains,
humans can retain responsibility.
And that is precisely
what AI should be used for.
Related repositories
The design approach described in this article —
AI systems that do not erase discontinuities, but handle judgment while keeping them explicitly separated —
is published as a set of design notes below.
These are not presented as finished products.
Their purpose is to preserve, as structure:
-
whose judgment diverged,
-
at which point,
-
and why the process stopped.
- multi-agent-orchestration-design
Readable and auditable multi-agent process orchestration
Contract-first AI decision system design to prevent PoC failure
Multi-agent system design treating local currency as distributed decision infrastructure
These are not systems designed to produce consensus.
They are designs intended to accurately record the fact that agreement was not reached — and that the process stopped there.
コメント