Agentic Orchestration of Multi-Provider Video Generation Pipelines: Automating Real Estate Marketing via Claude Code, Hexfield AI, and Vell AI
The paradigm of digital content creation is shifting from manual, tool-centric editing to agentic orchestration. Traditionally, generating high-fidelity marketing assets—such as cinematic property reels or animated digital brochures—required a fragmented workflow involving web scraping for data extraction, manual image selection, frame-by-frame video generation, and complex post-production in software like Adobe Premiere Pro or After Effects.
However, by leveraging Claude Code as an orchestration layer, we can now implement a provider-agnostic pipeline that automates the entire lifecycle of content production: from URL ingestion to final render. This post explores the architecture of a system capable of transforming a single real estate listing URL into professional-grade marketing assets using agentic workflows and multi-model integration.
The Architecture: Claude Code as the Orchestration Layer
The core innovation in this workflow is treating an LLM not merely as a conversational interface, but as a software engineer managing a complex state machine. In this architecture, Claude Code serves as the orchestration layer. It does not simply "write code"; it executes a predefined Technical Plan and utilizes custom Skills to interact with external APIs and CLI tools.
The workflow follows a structured design pattern:
- Data Extraction: Utilizing Playwright to scrape property listing data, including high-resolution imagery and metadata (price, location, features).
- Asset Selection Logic: An agentic decision-making process to identify the most "cinematic" exterior and interior shots based on visual composition.
- Video Generation Pipeline: Dispatching selected frames to a video generation provider.
- Assembly & Motion Design: Using HyperFrames to stitch clips, apply branded overlays (typography, color palettes), and implement camera movements.
By providing Claude Code with a rigorous technical plan, we minimize the "hallucative" tendencies of LLMs. We instruct the agent to follow a specific implementation roadmap, document architectural decisions, and—crucially—to halt execution and query the user when parameters like branding or music assets are undefined.
Pipeline Implementation: Hexfield AI Integration
The first iteration of our pipeline utilizes Hexfield AI. This implementation relies on the hexfield-ai CLI and a custom-built "Skill" integrated directly into the Claude Code environment.
Setup and Authentication
The integration begins with the installation of the Hexfield CLI within the development environment, followed by an authentication handshake (hexfield auth) to establish secure API communication. The deployment of a specialized Hexfield Skill allows Claude Code to trigger video generation tasks programmatically.
Model Configuration: Google VO3 and GPT-driven Image Handling
For the underlying generative heavy lifting, we configure the pipeline to utilize Google’s VO3 model as the primary video diffusion engine. This model is tasked with transforming static 2D images into cinematic clips characterized by realistic camera pans, tilts, and zooms. To supplement this, GPT-based models are utilized for image handling tasks—specifically for analyzing metadata and performing any necessary generative image augmentation required to maintain aspect ratio consistency during the transition from static photo to video frame.
Achieving Provider Agnosticism: The Vell AI Swap
A critical requirement for a robust production pipeline is the ability to swap underlying inference providers without refactoring the entire orchestration logic. This is achieved by decoupling the Creative Direction Prompt (the "what") from the Provider Implementation Skill (the "how").
In our second implementation, we replaced Hexfield AI with Vell AI, a platform providing an abstracted API layer for accessing a diverse collection of generative models.
The Swap Logic
The transition involved:
- Updating the environment variables with a new
VELL_AI_API_KEY. - Adapting the custom skill to interface with Vell AI’s specific endpoint architecture.
- Maintaining the exact same Technical Plan and Prompt Library.
Because the high-level prompt defines only the creative outcomes (pacing, motion design, visual style), Claude Code can execute the same instructions against a different model backbone. This allows developers to perform side-by-side A/B testing of video quality and cost-efficiency without rewriting the automation logic.
Output Analysis: Social Media Reels vs. Animated Digital Brochures
The pipeline is designed to produce two distinct classes of marketing assets, each requiring unique motion design parameters within the prompt library.
1. The Social Media Reel
The objective here is high-impact, short-form engagement. The prompts instruct the agent to focus on rapid pacing, aggressive transitions, and high-frequency cuts. The goal is not professional-grade color grading but rather "good enough" quality that maintains low production latency and extremely low cost-per-asset.
2. The Animated Digital Brochure
Conversely, the digital brochure requires a premium, luxury aesthetic. The prompt architecture for this asset shifts focus toward:
- Canvas Structure: Defining complex layouts and page transitions.
- Motion Design: Implementing smooth, slow-motion camera movements that mimic a "living" presentation board.
- Typography & Branding: Integrating intentional typographic overlays and clean design systems to elevate the perceived value of the property.
Conclusion: The Workflow as the Product
The ultimate takeaway from this implementation is that the value proposition has shifted. In an era of ubiquitous generative models, the specific model (Hexfield vs. Vell AI) becomes secondary to the orchestration workflow.
By packaging data extraction, asset selection, multi-model generation, and automated assembly into a single, repeatable agentic loop, we have created a scalable content engine. This architecture is fundamentally modular; while demonstrated here for real estate, the same orchestration logic can be applied to e-commerce product advertising, travel marketing, or any domain requiring high-volume, automated visual content production.