NousResearch/kanban-video-pipeline

NousResearch/kanban-video-pipeline

Language: Python

Stars: 34

Forks: 7

Open issues: 0

Created: 2026-05-02T17:01:19Z

Pushed: 2026-05-02T17:01:23Z

Default branch: main

Fork: no

Archived: no

README:

Kanban Video Pipeline — Multi-Agent Creative Production

> A recursive demonstration: the Hermes Agent Kanban system creates a video explaining itself.

Four AI agent profiles collaborate autonomously — a director, cinematographer, renderer, and editor — communicating through structured handoffs, comments, and shared workspace files. One command kicks off the entire pipeline; you come back to a finished video.

What This Is

A Kanban-driven multi-agent video production pipeline built on Hermes Agent. You describe a video, the system decomposes it into a task graph, and specialized agent profiles execute each stage:

You type one command
→ Director decomposes the brief into 9 tasks
→ Cinematographer designs visual language
→ 7 Renderers execute in parallel (ASCII video + p5.js)
→ Editor reviews clips + assembles final cut
→ output/final.mp4

The entire pipeline auto-runs with zero human intervention. Every stage promotes automatically when its dependencies complete.

The Recursive Concept

This video is about Kanban — and it was made by Kanban. The first project demonstrates the system by having it create an explainer video about itself. Two-part structure:

1. Part 1 (captured with Screen Studio): You type the command, the dashboard lights up, tasks flow between columns 2. Part 2 (generated by the pipeline): The actual video the system produced — ASCII art + generative visuals explaining how Kanban works

The Team

| Profile | Role | Communication | |---------|------|---------------| | Director | Vision, decomposition, final approval | Handoffs to cinematographer, comments on anything | | Cinematographer | Visual spec, quality gate | Specs to renderers, feedback via comments | | Renderer | Code execution (ascii-video / p5.js) | Reads parent handoff + comments, writes clips | | Editor | Pacing review, assembly, final cut | Reviews with AI video analysis, assembles with ffmpeg |

Cross-Profile Communication

Agents talk to each other through three built-in Kanban channels:

• Structured handoffs — when a task completes with summary + metadata, every child task automatically receives it in context
• Comments — any profile can write kanban_comment(task_id=) on any task; the next worker reads the full thread
• Shared workspace — all profiles read/write to the same project directory

Quick Start

Prerequisites

• Hermes Agent installed
• OpenRouter API key (for claude-opus-4.6)
• Gateway running (hermes gateway start)

Setup

chmod +x setup.sh
./setup.sh

This creates 4 profiles (director, cinematographer, renderer, editor), writes their personalities, sets up the project structure, and initializes the Kanban board.

Run

hermes kanban create "Make a video explaining how Hermes Kanban works" \
--assignee director \
--workspace dir:$(pwd)/kanban-launch-video \
--tenant video-pipeline \
--priority 2 \
--max-runtime 4h \
--body "Read brief.md. Decompose into: cinematographer → renderers → editor. All child tasks MUST use workspace_kind='dir' and workspace_path='$(pwd)/kanban-launch-video'. Pass tenant='video-pipeline' on every create."

Monitor

hermes kanban watch --tenant video-pipeline # live events
hermes kanban list --tenant video-pipeline # board state
hermes dashboard # visual board (Kanban tab)

Project Structure

├── README.md ← you are here
├── ARCHITECTURE.md ← system design, communication model
├── RUN.md ← step-by-step execution guide
├── brief.md ← the video specification
├── setup.sh ← one-shot setup script
├── taste/
│ ├── brand-guide.md ← visual language rules
│ └── emotional-dna.md ← the creative soul document
├── tools/
│ └── media_analyze.py ← AI audio/video analysis CLI (Gemini via OpenRouter)
└── render-scripts/ ← what the agents actually wrote (reference)
├── scene-01/render.py ← ASCII video (Python + NumPy + ffmpeg)
├── scene-02/render.py
├── scene-03/sketch.html ← p5.js generative (HTML + Puppeteer)
├── scene-04/render.py
├── scene-05/sketch.html
├── scene-06/render.py
└── scene-07/render.py

Output

The pipeline produces:

• output/final.mp4 — 1920×1080 @ 24fps, ~2 minutes, with audio
• output/final-noaudio.mp4 — silent version for re-scoring
• Individual scene clips at scenes/scene-NN/clip.mp4

How It Works (Deep Dive)

See [ARCHITECTURE.md](./ARCHITECTURE.md) for:

• Full communication model diagram
• Task graph with dependency links
• Iteration loops (cinematographer ↔ renderer, editor ↔ renderer)
• Renderer flexibility (swap ascii-video / p5js / remotion per task)
• Media analysis tool usage

Timing

| Phase | Duration | What Happens | |-------|----------|--------------| | Director | ~3 min | Reads brief, creates 9 tasks | | Cinematographer | ~8 min | Designs visual spec for all scenes | | Renderers (×7 parallel) | ~20 min | ASCII/p5.js rendering | | Editor | ~10 min | Reviews, assembles, muxes audio | | Total | ~45 min | Fully autonomous |

Adapting This

To make your own video: 1. Edit brief.md with your video spec 2. Replace taste/brand-guide.md with your visual language 3. Drop your audio into the workspace 4. Run setup.sh → fire the kanban command 5. Come back to a video

The renderer profile has both ascii-video and p5js skills — use --skill on individual tasks to control which renderer is used per scene.

License

MIT