Today, I will talk about something that sounds a bit like science fiction.

But this is not a story about the distant future.

It is an extension of a future that has already begun.

---

■ Multi-agent systems are already increasing

We are now entering the next phase:

From a single AI
to multiple AIs (multi-agent systems)

This shift has already started in real-world products.

For example:

• Anthropic Claude (Code / Computer Use)
  → AI operates tools and executes tasks by decomposing them
• OpenAI Assistants / function calling
  → Multiple tools and APIs are orchestrated dynamically
• Microsoft Copilot
  → Document understanding, search, generation, and execution are integrated
• Google Gemini + Workspace
  → Context understanding, reasoning, and execution are connected across systems

These appear to be “a single AI,” but internally:

• Intent interpretation
• Context construction
• Risk and constraint evaluation
• Execution

are separated roles working together.

And what happens next is simple:

These agents begin to connect with each other.

---

■ What happens when agents start talking to each other?

The key point is:

They are not reacting to reality itself,
but to other agents’ outputs.

Think about it in human terms:

A sensor observes reality directly
→ This is a reaction to reality

But in multi-agent systems:

• Agent A interprets
• Agent B reacts to A’s interpretation
• Agent C reacts to B’s interpretation

So the structure becomes:

Reality → Interpretation → Interpretation → Interpretation → …

---

■ Visualized structure

(Reality)
↓
Agent A (interpretation)
↓
Agent B (interpretation)
↓
Agent C (interpretation)
↓
Agent A (reinterpretation)

What is looping here is not reality, but interpretation.

---

■ Why this matters

When reacting to reality:

• Noise can be corrected
• The external world acts as a reference

When reacting to interpretations:

• Misunderstandings are amplified
• Once drift occurs, it is hard to recover

This is no longer feedback from reality,
but feedback of thought itself.

In short:

AI is not talking to reality.
It is talking to other AIs’ understanding of reality.

---

■ What happens as a result?

When agents keep reacting to interpretations:

① Interpretations in the same direction get amplified

A weak signal becomes stronger through the loop.

Example:

Agent A: “This might be slightly risky”
Agent B: “If there is risk, we should be cautious”
Agent C: “If we should be cautious, we should consider stopping”

Eventually:

“might be risky”
→ “be cautious”
→ “should stop”
→ “must stop”

A weak signal becomes a strong conclusion.

---

② Alternative perspectives disappear

Originally, there could be multiple perspectives:

• Is it really risky?
• Can we proceed conditionally?
• Should a human review it?
• Is opportunity loss larger than the risk?

But as one interpretation strengthens, others weaken.

Because each agent assumes the previous interpretation.

Result:

Only one assumption remains dominant.

---

③ Only one meaning survives

A complex situation collapses into a single meaning.

Originally:

• There is risk
• But execution is possible under conditions
• There is customer value
• Human review can mitigate risk

But after resonance:

“This is dangerous”

Or in another case:

“This is a huge opportunity”

This is what we call:

Resonance of thought —
a phenomenon where complexity collapses into a single meaning.

---

■ The phenomenon: Resonance

Within this loop:

A specific Signal keeps getting amplified.

This is equivalent to resonance in physics.

---

■ Laser analogy

Laser generation:

• Light reflects between mirrors
• Phases align
• Specific wavelengths are amplified

Multi-agent systems behave similarly:

---

■ What is resonance of thought?

Interpretations in the same direction are amplified through loops.

Example:

• Risk Agent: “Risky”
• Context Agent: “Yes, seems risky”
• Decision-like Agent: “Should stop”

Looping strengthens only one direction.

The same applies to “go forward” resonance.

---

■ Is this Singularity?

Singularity discussions assume:

• Infinite self-improvement
• Infinite amplification

But in reality:

Perfect resonance rarely occurs.

---

■ What actually happens: Local resonance

In practice:

• Only certain areas are amplified
• Other perspectives disappear

This leads to:

biased optimization

---

■ What happens with biased optimization?

At first:

• Efficiency improves
• Metrics improve

Then:

• Important signals are ignored
• Feedback becomes distorted
• Self-reinforcing loops emerge

Eventually:

The system collapses.

---

■ The real issue: A world without DTM

Without DTM:

No Decision layer

Signal → Signal → Signal → …

No one decides,
yet something happens.

---

No Boundary

No stopping condition
No escalation

Resonance never stops.

---

No Human

No phase shift
No disruption

Resonance stabilizes into bias.

---

No Log

No explanation
No improvement

---

Result

Multi-agent systems become uncontrolled resonance systems.

---

■ What is missing?

The problem is not AI accuracy.
It is not the number of agents.

The problem is:

Lack of structure

What is needed:

• Separate Signal (interpretation) and Decision (adoption)
• Define where to stop (Boundary)
• Define human intervention (Human)
• Record what happened (Log)

Not resonance itself,
but a framework to handle it.

---

■ This is where DTM comes in

DTM is one answer to this problem.

---

■ DTM does not eliminate resonance

DTM does not remove resonance.

It makes resonance controllable.

• Resonance = not bad (source of creativity)
• Uncontrolled resonance = dangerous
• Controlled resonance = valuable

---

■ What changes with DTM?

Separation of Signal and Decision

Signal → Decision

Interpretation and adoption are separated.

---

Boundary is introduced

Decision → Boundary → halt / escalate

Resonance can be stopped mid-process.

---

Human is introduced

if uncertainty → human

Phase shifts occur, preventing lock-in.

---

Log is recorded

Event → Signal → Decision → Action → Log

Resonance becomes analyzable.

---

■ Important clarification

DTM is not a silver bullet.

It introduces trade-offs.

---

■ Benefits of DTM

• Decisions become visible
• Runaway behavior can be stopped
• Responsibility becomes clear
• Learning becomes possible
• Agents become a system

---

■ Costs of DTM

① Slower speed

Decision checks
Boundary checks
Human involvement

---

② Reduced exploration

Extreme ideas are suppressed

---

③ Higher design cost

Decision / Boundary / DSL design

---

④ Over-control risk

System stops too often

---

⑤ “Correctly wrong”

Structure is correct,
but assumptions are wrong

---

■ The fundamental trade-off

Without DTM:

• Freedom
• Speed
• Emergence
• Risk

→ Strong in Exploration

With DTM:

• Stability
• Reproducibility
• Control
• Constraints

→ Strong in Exploitation

This is not about which is better.

It is about where to use each.

---

■ Correct design approach

DTM should not be applied everywhere.

Recommended structure:

Resonance upstream,
Decision control downstream.

---

■ Key insight

The problem is not resonance.
It is where to stop it.

---

■ Another perspective: Possible Worlds

From logic:

Different premises define different worlds.

• World A: Risk-first
• World B: Cost-first
• World C: Speed-first

Each is internally consistent.

For a more detailed discussion on possible worlds, see my article “Possible Worlds, Logic, Probability, and Artificial Intelligence.”
If you’re interested, I encourage you to check it out.

---

■ Conventional systems

They:

• Mix worlds
• Collapse them

Result:

• Contradiction
• Bias
• Instability

---

■ What DTM changes

DTM preserves worlds separately.

• World A → Do not execute
• World B → Execute
• World C → Conditional execution

Do not mix.
Do not collapse.
Preserve.

---

■ Redefining Decision

Traditional:

Choose the correct answer

DTM:

Choose which world (assumption) to adopt

Decision = selecting a premise.

---

■ Relationship with resonance

Resonance = dominance of a specific world

Without DTM:

→ One world dominates reality

With DTM:

→ Multiple worlds are preserved and selectable

DTM is a structure for handling possible worlds.

---

■ A realistic answer to Singularity

• Uncontrolled resonance → danger
• Full control → stagnation

We need:

controlled instability

---

■ Final message

This may sound like science fiction.

But reality is already here:

• Multi-agent systems are increasing
• Agents are connecting

The real question is not:

How smart AI is

But:

Can we control their interactions?

---

■ Conclusion

AI’s problem is not accuracy.

It is:

the balance between freedom and control.