Skip to content

Craftology Architectural Overview

Craftology is a next-generation Creative Intelligence Framework — an AI-native operating system built to automate and streamline content production. At its core, the system is organized into four main components.

--- config: layout: dagre --- flowchart TB subgraph UI_Layer["UI / Interaction Layer"] User["User / Creator"] Chat["Chat Interface"] end subgraph Orchestration_Layer["Agent Orchestration Layer (Behavior)"] direction TB Orchestrator["AI Producer (Master Orchestrator)"] Planner["Golden Path Engine (Planner & Router)"] Specialist_Agents["Specialist_Agents"] end subgraph Specialist_Agents["Reactive Specialist Agents"] direction TB StoryCraft["StoryCraft Agent"] AssetProd["Asset Production Agent"] OpsAgent["Ops/Budget Agent"] GovAgent["Governance/Policy Agent"] end subgraph Intelligence_Layer["Business Logic & Intelligence"] WorldModel["Unified Project Graph (World Model)"] Context["Context"] Lineage["Lineage"] end subgraph Resource_Pools["Model / Resource Pools"] ExternalAPIs["External APIs (OpenAI, Anthropic, Sora)"] HostedPools["Hosted Pools (Llama, Stable Diffusion)"] MCP["MCP Servers & Tools"] end subgraph Infrastructure_Layer["Infrastructure & Model Routing"] ModelSelector["Global Model Selector"] Resource_Pools end User -- Provides Input --> Chat Chat -- Sends Intent --> Orchestrator Orchestrator -- Decomposes Intent --> Planner Orchestrator -- Routes Task --> Specialist_Agents Specialist_Agents -- Perceives & Updates --> WorldModel WorldModel --> Context & Lineage Specialist_Agents -- Executes via --> ModelSelector ModelSelector --> ExternalAPIs & HostedPools & MCP Chat -- Display Result --> User Planner@{ shape: hex} WorldModel@{ shape: cyl} Planner:::yellow WorldModel:::lightblue classDef yellow stroke:yellow,stroke-width:2px classDef lightblue stroke:blue,stroke-width:2px

Intelligence Substrate: Unified Project Graph & World Models

  • Unified Project Graph: This is the heart of Craftology, storing story structures, character arcs, asset metadata, and project rules. It goes beyond a traditional file system, serving as a single source of truth for the entire project.
  • World Model: Built on the Project Graph, the World Model tracks state and causality across production modules like Storycraft, Assets, Shots, and Editorial. Future updates will enable runtime learning, persona consistency, and multi-agent orchestration in 3D and video environments.

Agentic Framework: Reactive & Specialist Agents

Craftology runs a network of self-learning agents that handle tasks across the production pipeline:

  • Reactive Agents: Continuously monitor project state, predict outcomes, and choose the best actions based on data, user feedback, and constraints.
  • Specialist Agents: Bring deep expertise in specific areas:
    • StoryCraft Agents: Ensure narrative quality through script analysis and beat structuring.
    • Production Agents: Handle asset creation, style transfer, and composition.
    • Operational Agents: Manage budgets, schedules, and resource optimization.
    • Governance Agents: Oversee policies, safety, and data lineage.

Orchestrator: AI Producer & Director-Level Reasoning

  • AI Producer (Global Selector Agent): Serves as the central planner, applying director-level reasoning to guide the entire project:

    • Long-Horizon Planning: Creates the "Golden Path" from concept to final release.
    • Cross-Episode Continuity: Maintains consistency across workflows.
    • Multi-Step Orchestration: Coordinates sequences like Generate → Critique → Improve → Validate → Publish.
    • Dynamic Optimization: Chooses the most efficient agents and models to save time and cost.

  • Golden Path Engine: Automatically generates and updates the optimal production roadmap, adjusting to edits, failures, or new constraints.

Technical Architecture & Infrastructure

--- config: theme: redux-color --- sequenceDiagram autonumber participant User as User (Director) participant Prod as AI Producer (Orchestrator) participant GPE as Golden Path Engine<br/> (Planner & Reasoner) participant Agent as Reactive Agent participant Graph as Unified Graph<br/>(World Model) participant Infra as Model Infrastructure %% ------------------------------- %% Phase 1: Planning %% ------------------------------- Note over User,Prod: PHASE 1 — Planning & Decomposing the Intent User->>Prod: Creative Intent<br/>("Make scene darker") Prod->>GPE: Request Optimal Workflow Plan GPE->>Graph: Query Project State & Constraints<br/>("Current Context") Graph-->>GPE: Return Current Context GPE->>GPE: Planning & Reasoning<br/>Decompose into Task Chain GPE-->>Prod: Return "Golden Path Plan" %% ------------------------------- %% Phase 2: Execution Loop %% ------------------------------- loop For each task in the plan Note over Prod,Agent: PHASE 2 — Task Execution & Reactive Behavior Prod->>Agent: Delegate Task<br/>(including context) Agent->>Agent: Reactive Behavior Check<br/>("What is happening now?") Agent->>Infra: Select Best Model & Execute Infra-->>Agent: Raw Output Agent->>Agent: Self-Critique / Validate Output Agent-->>Prod: Task Complete / Review Ready Prod->>User: Propose Structured Options<br/>(HITL Review) end alt User Approves User-->>Prod: Approve Prod->>Graph: Update Unified Graph State else User Rejects / Adjusts User-->>Prod: Feedback / Adjustment Prod->>GPE: Re-plan / Re-route<br/>(Dynamic Optimization) GPE->>Graph: Request Updated Context Graph-->>GPE: Updated Context GPE-->>Prod: Updated Plan / Route end Note over User,Infra: Continuous loop of plan → execute → validate → HITL → update or re-plan
  • Model-Agnostic: Tasks are assigned by the Global Selector Agent to the most suitable models—whether OpenAI/Anthropic for narrative, Latent Diffusion/Sora/Veo for asset creation, or local/hosted models for private workflows.
  • Human-in-the-Loop (HITL): The architecture includes HITL as a first-class mechanism.
  • Embedded Governance: Built-in systems like the Ops Hub manage policies and compute, IP-Safe Zones protect sensitive data, and audit trails ensure compliance.
  • Agent Improvement Service (AIS): Continuously learns from agent activity and user interactions, automating evaluations and refining the Golden Path for ongoing improvement.