Behavior Tree × LLM × Knowledge Graph — A Three-Layer Architecture for AI Decision Systems in Multi-Agent Orchestrators —

AI systems have evolved rapidly in recent years.

However, most systems are still designed around the following structure:

• Models
• APIs
• Data

With this structure, it is difficult to fully represent:

• The structure of AI decisions
• The behavior of multi-agent systems

Structuring Decisions

To properly structure decision-making, three key technologies become essential:

• Behavior Tree
• LLM (Large Language Model)
• Knowledge Graph

By combining these, we can build:

👉 Explainable and controllable AI decision systems

The Three Roles of AI Decision Systems

An AI decision system can be decomposed into three major roles:

1. Knowledge
   The semantic structure of the world
2. Reasoning
   Understanding based on knowledge
3. Execution
   Control and realization of decisions

These roles are handled by:

• Knowledge Graph → Knowledge (Meaning)
• LLM → Reasoning
• Behavior Tree → Execution Control

Knowledge Graph (Meaning Structure)

A Knowledge Graph represents:

👉 The semantic structure of the world

For example:

User
 ├ segment
 ├ purchase_history
 ├ risk_score

This can be interpreted as:

• User: Data representing user attributes and state
• Offer: Actions or conditions presented to the user

It also expresses relationships:

User → belongs_to → Segment
Offer → part_of → Campaign
Campaign → has_budget → Budget

This means:

👉 Users belong to segments, offers are part of campaigns, and campaigns are constrained by budgets.

Through this structure, AI can understand:

👉 The meaning of the world

LLM (Reasoning)

LLMs perform:

👉 Semantic reasoning

Example:

User input:

“I want to cancel my subscription”

LLM interprets this as:

intent = cancel_subscription

LLMs are also used for:

• Summarization
• Knowledge inference
• Document understanding
• Rule interpretation

However:

👉 LLMs cannot manage the execution structure of decisions

That role belongs to the Behavior Tree.

Behavior Tree (Decision Structure)

A Behavior Tree represents:

👉 The execution structure of decisions

Example:

Selector
 ├ RiskCheck
 ├ BudgetCheck
 ├ PolicyCheck
 └ ExecuteOffer

This means:

👉 The system checks risk, budget, and policy in order, and executes the offer if all conditions are satisfied.

Behavior Trees naturally express:

• Conditional branching
• Priority handling
• Fallback logic

Originally used in:

👉 Game AI

But they are highly suitable for:

👉 Controlling multi-agent AI systems

The Three-Layer Architecture

Combining:

Behavior Tree × LLM × Knowledge Graph

Results in the following structure:

Knowledge Graph
 ↓
LLM
 ↓
Behavior Tree

Roles:

This forms:

👉 The fundamental architecture of AI decision systems

Integration with AI Orchestrators

This three-layer structure operates within an AI orchestrator:

Event
 ↓
Knowledge Graph Query
 ↓
LLM Reasoning
 ↓
Behavior Tree Decision
 ↓
Policy Check
 ↓
Boundary
 ↓
Execution

This represents:

👉 A decision-making process where an event triggers knowledge retrieval, reasoning, structured decision-making, constraint validation, and finally execution.

This enables separation of AI decisions into:

• Meaning
• Reasoning
• Execution

Relationship with Decision Trace

In the Decision Trace Model:

AI does not only produce results.

It records:

👉 How the decision was made

The trace includes:

• Event: What happened
• Signal: What AI recognized
• Decision: What was chosen
• Policy: Applied rules
• Boundary: Constraints
• Execution: Final action

Internal Roles

Within this process:

• Knowledge Graph → Provides meaning
  Defines relationships and context
• LLM → Performs reasoning
  Interprets situations and generates candidates
• Behavior Tree → Controls decision flow
  Selects decisions based on conditions and priorities

Core Insight

👉 Not only the result, but the entire process
(Meaning → Reasoning → Decision Flow)
is recorded as a trace.

In other words:

👉 It records the structure of decisions, not just the outcomes

Why These Three Technologies Matter

Most modern AI systems are:

👉 LLM-centric

But LLMs alone cannot provide:

• Reproducibility
• Controllability
• Auditability

That is why the three-layer structure is essential:

• Knowledge Graph (Meaning)
• LLM (Reasoning)
• Behavior Tree (Execution)

This transforms AI from:

👉 A black box

into:

👉 A structured decision system

The Future of AI: Decision Architecture

The evolution of AI is not determined by:

👉 Model size

The real differentiator is:

👉 Decision architecture

Conclusion

• Behavior Tree
• LLM
• Knowledge Graph

These three technologies are the key to transforming AI into:

👉 Controllable and auditable decision systems

And at the center of this transformation lies:

👉 The AI Orchestrator

Chinoba

Chinoba

Intelligence as Relationship

Research Platform

founded by

Masao Watanabe

AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance