Workflow Integration

Data Flow and Execution Pipeline

The workflow is a linear pipeline with recursive feedback loops, designed to move from raw text to structured video directives.

Integration Flow

The following steps outline the transformation of a raw text chapter into a series of production-ready video generation prompts and metadata.

Step 1: The Profiler (Pre-processing)

The first stage builds the “Story Bible.” An agent rips through the full text to identify every recurring entity and location.

1

The Profiler

Pre-processing Phase
Input
Full Book Text
Action
Identify Characters & Locations
Output
characters.jsonlocations.json

Step 2: The Chunking Engine

Once assets are defined, the chunking engine segments the text into 8-10 second beats, weighting them based on narrative density.

2

The Chunking Engine

Segmentation Phase
Input
Chapter_01.md
Action
Time-Weighted Segmentation
Output
Chapter_01_chunks.json

Step 3: The Continuum Flow

This is the core context enrichment cycle. For every chunk, the agent queries the “Backbone” to inject character details and state information.

3

The Continuum Flow

Context Enrichment
Input
Chunk N + Context Backbone
Action
Enrich with visual descriptors
Output
Enriched Prompt N

Step 4: The Director Agent

The Director translates narrative prose into technical camera direction (Shot types, angles, lens info).

4

The Director Agent

Technical Translation
Input
Enriched Prompt N
Action
Generate Camera Directives
Output
directives/Scene_01_Seg_N.json

Step 5: Context Update

Finally, the system compresses the events of the current chunk to update the Level 1/2 summaries for future retrievals.

5

Context Update

Closing the Loop
Input
Chunk N Narrative
Action
Recursive Summarization
Output
Updated Backbone

Agentic Workflow Expansion: The “Production Studio” Model

🧠
Global Orchestrator

The Showrunner

The “Brain” that manages priorities. Pauses production if new context (e.g., a new character) is discovered.

🛑
Feedback Loop

The QA Critic

The “Editor” that rejects bad prompts. Enforces strict character consistency before rendering.

📝
Context Swarm

Writers’ Room

Parallel agents mining the text for “Story Bible” data. Resolves cross-chapter conflicts.

🎨
Visual Assets

Art Department

Casting Agents & Location Scouts generating reference images and LoRAs asynchronously.

Our initial architecture defined a linear pipeline—a conveyor belt. While efficient, it lacks the resilience of a true Agentic AI system. In a real-world deployment, the system must handle ambiguity, errors, and parallel tasks without human intervention. The “Missing Link” is The Feedback Loop (The Critic) and Asynchronous Orchestration.

2. The Core Agentic Roles

Instead of a single “Process,” we define distinct autonomous agents functioning as a digital film crew.

A. The Showrunner (Global Orchestrator)

Role: The “Brain” of the operation.

  • Agentic Action: Scans the book. Assigns “Job Tickets” to other agents.
  • Crucial Capability: Dynamic Re-prioritization. If Chapter 3 reveals a new main character, it pauses the Scene Generators.

B. The Art Department (Parallel Pre-Production)

Role: Visual Asset Generators.

  • Workflow: Casting Agent generates Reference Images; Location Scout generates Environment LoRAs.

C. The Writers’ Room (Context Swarm)

Role: The “Continuum Flow” Implementation.

  • Agentic Action: Parallel summarization with cross-chapter reconciliation.

D. The QA Critic (The Feedback Loop)

Role: The “Editor.”

  • Workflow: Rejecting bad prompts and forcing retries until output matches character state.

4. Summary of Improvements

FeatureLinear Pipeline (Old)Agentic Swarm (New)
ProcessingSequential (Slow)Asynchronous / Parallel (Fast)
Error HandlingFails at end of pipeSelf-corrects mid-stream (Critic)
ContextPassive RetrievalActive Reconciliation (Showrunner)
CostWasted on bad promptsSaved by QA rejection before render