Graph-based intermediate representation optimized for AI comprehension.
Verifiable semantics, native performance, and human-in-the-loop validation.
Traditional development: AI assists humans in writing human-readable code
NAL: Humans describe intent, AI agents write machine-optimized graphs
| Traditional Workflow | → | NAL Workflow |
|---|---|---|
| Human writes C++/C# code | → | Human writes natural language intent |
| Compiler reads text → AST | → | AI agents generate semantic graphs |
| Manual debugging & testing | → | Human verifies intent in semantic IDE |
| Code must be human-readable | → | Graphs optimized for AI & compiler |
| Hours to days iteration | → | Faster iteration cycles |
JSON/JSONL serialized graphs representing game logic. First-class support for ECS, physics, and shaders.
Core primitives: Entity, Component, System, PhysicsDomain, MaterialGraph. Declarative rather than imperative.
Specialist fine-tuned SLMs collaborate: ECS Architect, Physics Agent, Shader Agent, Orchestrator.
Every node includes metadata: created_by_agent, source_intent_id, type annotations. Full lineage tracking.
Stack-based virtual machine for hot reload support. Executes NAL in <1 second iterations. JIT compilation via LLVM/Cranelift planned for Phase 4.
Designed for real games, not demos. Planned UE5 integration for AAA workflows.
NAL is NOT human-readable code. It's intentionally opaque to maximize semantic density.
Humans become "semantic engineers" verifying intent, not debugging code. Different role, not eliminated.
NAL is game-dev specialized. Optimized for ECS patterns, not web apps or databases.
Unique node ID representing a game object. Lightweight identifier with no data itself.
{
"entity_id": "player_001",
"archetype": "PlayerCharacter"
}
Pure data structures attached to entities. No logic, only state.
{
"Position": {
"x": "float",
"y": "float"
}
}
Pure functions querying and transforming component data. Auto-parallelizable.
{
"system": "ApplyVelocity",
"query": ["Position", "Velocity"],
"transform": "..."
}
Declarative physics properties: constraints, materials, collision rules.
{
"physics": {
"gravity": -9.8,
"friction": 0.3
}
}
Shader graph definitions compiled to HLSL/MSL/SPIR-V.
{
"material": "PBR",
"nodes": [...]
}
Every node tracks creation agent, source intent, and type information.
{
"created_by_agent": "ECS_Architect",
"source_intent_id": "player_v1"
}
A simple player controller in NAL graph format
{
"metadata": {
"name": "PlayerController",
"version": "1.0.0",
"created_by_agent": "ECS_Architect",
"source_intent_id": "player_movement_basic",
"timestamp": "2025-12-28T00:00:00Z"
},
"components": {
"Position": {
"x": { "type": "float", "default": 0.0 },
"y": { "type": "float", "default": 0.0 }
},
"Velocity": {
"dx": { "type": "float", "default": 0.0 },
"dy": { "type": "float", "default": 0.0 }
},
"Player": {
"speed": { "type": "float", "default": 200.0 },
"jump_force": { "type": "float", "default": 400.0 }
}
},
"systems": {
"ApplyVelocity": {
"query": ["Position", "Velocity"],
"execution_order": 10,
"transform": {
"Position.x": "Position.x + Velocity.dx * DeltaTime",
"Position.y": "Position.y + Velocity.dy * DeltaTime"
},
"metadata": {
"created_by_agent": "ECS_Architect",
"source_intent_id": "basic_movement"
}
},
"PlayerInput": {
"query": ["Player", "Velocity"],
"execution_order": 5,
"inputs": ["keyboard"],
"transform": {
"Velocity.dx": "if KeyPressed(Left) then -Player.speed else if KeyPressed(Right) then Player.speed else 0",
"Velocity.dy": "if KeyPressed(Jump) and OnGround then Player.jump_force else Velocity.dy"
},
"metadata": {
"created_by_agent": "ECS_Architect",
"source_intent_id": "player_input_handling"
}
}
},
"entities": [
{
"id": "player_001",
"components": ["Position", "Velocity", "Player"],
"metadata": {
"archetype": "PlayerCharacter",
"created_by_agent": "ECS_Architect"
}
}
]
}
Every token carries maximal semantic information. AI agents can reason about entire game systems within LLM context windows.
Type systems and constraint checking at compile time. Catch bugs before runtime through static analysis.
Fine-tuned SLMs excel in their domain. ECS Architect understands entity patterns. Physics Agent knows constraints.
Semantic IDE shows intent graphs, not text. Humans verify "why" not "how". See which agent created each node.
Optimized interpreter runs games at 60fps. LLVM compilation in development for maximum throughput.
Intent → Graph → Compile → Play. Streamlined workflow with no manual coding or boilerplate.
Full lineage tracking from human intent to compiled code. Every decision traceable to its origin.
First-class ECS support. Compiler guarantees system purity, data access correctness, and automatic parallelization.
Describe physics properties, not implementation. Compiler chooses optimal physics engine and settings.
Specialist agents collaborate through A2A protocol
Role: Workflow management, task decomposition, agent coordination
Status: Planned for Phase 3 (fine-tuned SLM)
Role: Entity-Component-System design, archetype optimization
Status: Prototype complete (92% success on test cases)
Role: Physics simulation, constraint solving, material properties
Status: Planned for Phase 3 (fine-tuned SLM)
Role: Material graphs, shader compilation, visual effects
Status: Planned for Phase 3 (fine-tuned SLM)
Role: Game design validation, balance analysis, performance prediction
Status: Planned for Phase 3 (integrated in Harmonic IDE)
Role: Retrieval system for NAL patterns and game templates
Status: 93 examples indexed, expanding in Phase 2-3
| Feature | Traditional C++ | AI Copilot | NAL |
|---|---|---|---|
| Code Format | Text files (.cpp) | Text files (.cpp) | JSON graphs (.nal.json) |
| Readability | Human-readable | Human-readable | AI-optimized (opaque) |
| Authoring | Humans write code | AI assists humans | AI generates from intent |
| Debugging | Manual code debugging | Manual code debugging | Semantic verification |
| Performance | Native (C++ compiler) | Native (C++ compiler) | Native (Rust/LLVM) |
| Type Safety | Compile-time | Compile-time | Compile-time + metadata |
| Lineage Tracking | None | Comments only | Full metadata trail |
| ECS Support | Library-based | Library-based | Language primitive |
From JSON graphs to production-ready development environment
Focus: Stable Resonance Engine
Focus: NAL v2 Compiler + Harmonic Alpha
Focus: Fine-tuned SLMs + Harmonic Beta
Focus: Production-ready + Marketplace
Purpose-built IDE launching in Phase 2 (Months 5-8)
Fine-tuned SLMs run on your machine (no cloud, no API costs). Privacy-first, offline-capable. Generate NAL from natural language descriptions.
Edit NAL code, save, see changes in <1 second. No game restart. State preserved across reloads. 10x faster than Unity/Godot.
Blueprint-style visual scripting that generates NAL code. Drag-and-drop entities, components, systems. Round-trip: visual ↔ NAL.
Catches game design bugs ("Boss HP too high for player DPS"), not just code bugs. Balance analysis, pacing suggestions, performance predictions.
Indie: $19/mo (1 seat, basic AI)
Studio: $49/mo/seat (5+ seats, advanced AI)
Enterprise: Custom pricing
Try Harmonic IDE risk-free. No credit card required. Full feature access during trial period. Alpha users get lifetime discounts.
NAL represents a fundamental shift in how games are created.