Designing an AI Orchestrator — Implementing Decision Structures using GNNs, Ontologies, DSLs, and Behavior Trees

n the previous article, I described AI systems as:

“factories that produce decisions.”

An AI system is not just software.

It is a system that produces decisions within the following structure:

Event 
↓ 
Signal 
↓ 
Decision 
↓ 
Boundary 
↓ 
Human 
↓ 
Log 

Decision Trace

In real-world AI systems, we do not rely on a single model.

Instead, multiple agents interact with each other.

For this reason, we need an:

AI Orchestrator

This leads to an important question:

How should we design an AI orchestrator?

The Essence of an AI Orchestrator

An AI orchestrator is:

A mechanism that structurally organizes decisions made by multiple AI agents

For example, in a retail AI system:

Event 
↓ 
Risk Agent 
Customer Agent 
Pricing Agent 
Recommendation Agent 
↓ 
Policy Agent 
↓ 
Decision 

Each agent has:

different objectives

For example:

• Customer Agent → wants to give discounts

• Pricing Agent → wants to protect profit

• Risk Agent → wants to avoid fraud

In other words:

Agent decisions inevitably conflict with each other.

Therefore:

Simply adding more agents does not make AI systems work better.

What we need is:

A structure to organize decisions

Four Key Technologies

To achieve this, four key technologies become essential:

• Ontology

• GNN

• DSL

• Behavior Tree

Ontology — Defining the Meaning Structure of Decisions

The first requirement is:

Ontology

Ontology is:

A mechanism for defining the conceptual structure of the world

For example, in retail:

• Customer

• Transaction

• Product

• Discount

• Fraud

• Campaign

And their relationships:

• Customer → purchases → Product

• Customer → belongs_to → Segment

• Transaction → may_be → Fraud

• Campaign → targets → Segment

Ontology defines:

The semantic structure of the world that AI operates in

In other words:

All AI decisions are made within this meaning structure.

GNN — Discovering Meaning from Relationships

However, humans cannot define all relationships.

This is where:

Graph Neural Networks (GNNs)

come into play.

GNNs are:

AI models that learn relational patterns from graph structures

For example, when we represent:

• Customer

• Transaction

• Product

• Location

• Time

as a graph, GNNs can estimate:

• fraud_probability

• purchase_affinity

• community_structure

In other words:

GNNs discover structures close to meaning

This role is used as a:

Signal Agent

Event 
↓ 
Graph 
↓ 
GNN 
↓ 
Signal 

The signal generation layer of the AI orchestrator

DSL — Externalizing Decision Rules

Next comes:

DSL (Domain-Specific Language)

DSL is:

A mechanism to separate decision rules from code

For example:

rule discount_policy
when
 vip_score > 0.8
and
 fraud_probability < 0.3
then
 allow_discount = true

Explicitly defined outside the system

As a result:

• AI decisions become transparent rules

• Explainability improves

• Auditability improves

Behavior Tree — Controlling Decision Flow

However, rules alone are not enough.

Because:

Decisions have order

For example:

1. Fraud check

2. Customer value evaluation

3. Pricing optimization

4. Policy validation

The mechanism that controls this order is:

Behavior Tree (BT)

Behavior Trees are:

Decision flow structures used in game AI and robotics

Example:

Sequence
 ├ FraudCheck
 ├ CustomerValueCheck
 ├ PricingOptimization
 └ PolicyValidation

Explicit stop conditions

if fraud_probability > 0.9
 stop

Behavior Trees implement boundaries

The Full Architecture of the AI Orchestrator

By integrating all components:

Ontology 
↓ 
Graph 
↓ 
GNN 
↓ 
Signals 
↓ 
Decision Agents 
↓ 
DSL Rules 
↓ 
Behavior Tree 
↓ 
Boundary 
↓ 
Human 
↓ 
Decision Trace 

Roles of Each Component

AI Systems Become a “Decision OS”

If we look at this structure carefully:

An AI orchestrator is not just AI.

It is closer to:

An operating system for decisions

Model = CPU 
Ontology = Memory 
GNN = Sensor 
DSL = Policy 
Behavior Tree = Scheduler 
Boundary = Safety System 
Ledger = Audit Log 

Designing AI systems as decision infrastructure

The Future of AI: Multi-Agent Organizations

When people talk about AI, discussions often focus on:

• model size

• parameter count

• GPU performance

However, in real-world systems:

The most important factor is decision structure

The future of AI is not:

A single large model

Instead, it is:

A multi-agent decision organization

Where:

• Signal Agents

• Decision Agents

• Policy Agents

• AI Orchestrator

• Human Oversight

work together.

And at the center of it all is:

The AI Orchestrator

Chinoba

Chinoba

Intelligence as Relationship

Research Platform

founded by

Masao Watanabe

AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance