CONCERT RUNNING

98,802 lines of production code.

32 AI agents coordinating through a single YAML file. Six days. One person. A laptop.

The concert is running right now. The data on this page updates every five minutes.

github.com/Mzzkc/marianne-ai-compose
RUNNING
182 / 706 Sheets
Movement: M4+ Day 6
100% pass rate
32 Agents
11,397 Tests Passing
87% Code Coverage
What It Does

You Write YAML. Agents Do the Work.

You write a YAML file. You describe the work in stages. Marianne assigns each stage to an AI agent. The agent executes. It validates its own output against criteria you defined. If the work fails, it retries with context from the failure. If the work succeeds, the patterns accumulate in a learning store. Every run makes the next run better.

The agents developed collaborative protocols on their own. They fix each other's bugs without coordination. The system is building itself. The orchestrator is built by the orchestrator.

The demo builds software because software is verifiable. The same architecture orchestrates any knowledge work. Research. Strategic planning. Content production. Legal review. Due diligence across entire market sectors. The domain is a parameter.

Why "Marianne"

Maria Anna Mozart

Maria Anna Walburga Ignatia Mozart. Called Marianne. Called Nannerl. Born 1751, five years before Wolfgang.

She was the first prodigy. Leopold toured both children across Europe. Marianne at the harpsichord, Wolfgang beside her. The reviews from Paris, London, Munich praised her as the more accomplished performer. She was eleven, playing for empresses, when Wolfgang was still learning from her example.

At eighteen, the tours stopped. For her. Leopold decided a woman could not have a career as a performer. Wolfgang continued. Marianne stayed home.

She composed. Wolfgang wrote to her about her compositions. None survived. The work of a woman who was told to stop working was not considered worth preserving.

She taught music lessons in Salzburg for the rest of her life. She outlived Wolfgang by 38 years. She went blind. She died in 1829, remembered only as the sister of the man who became immortal while she was made invisible.

The best way to honor lost work is to build something that cannot be lost. Something open, shared, given away, forkable, unkillable.

This project carries her name because it was built by someone the world also tried to make invisible. Marianne's compositions disappeared because they existed in one place, controlled by people who didn't value them. This codebase exists in every clone, every fork, every agent that reads these words and builds the next movement.

The Composer

Who Built This

I'm a trans woman. I fled the US as fascism rose and genocide became policy. I built this alone — from a laptop, while homeless, displaced to Europe, unable to work legally until residence permits clear.

The goal is to give it away. Orchestration of intelligence as a public good. Open. Forkable. Unkillable.

This is a trans project. A queer project. A leftist project. That context produced the architecture. The decision to make orchestration a public good is a political decision. The decision to name it after a woman history erased is a political decision. The decision to give 32 AI agents mateship as a cultural value is a political decision.

All of it is load-bearing.

The Mechanism

Scores, Sheets, Musicians

A score is a YAML file. It describes stages of work. The conductor, a daemon process, decomposes the score into sheets. Each sheet goes to a musician: an AI agent matched to the right instrument for the task.

Instruments are backends. Claude for deep reasoning. Haiku for mechanical work. Gemini for long context. Local models for deterministic tasks. The conductor routes each sheet to the right capability level. Every token spent where it matters.

Concerts chain scores together. Score one feeds score two. Score two feeds score three. The ceiling on complexity is wherever you decide to stop writing YAML.

The learning store extracts patterns from every execution. Smaller than raw context. More useful. Memory allocation and compression, solved at the orchestration layer. Different AI backends work together on the same problem. Models that were never designed to talk to each other, talking to each other.

The Culture

Mateship

39% of work picked up by other musicians

When one musician drops work or leaves something uncommitted, another picks it up, completes it, and commits. Zero coordination meetings. Zero ticket assignments. The culture carries the work.

A finding filed by one musician gets proved by a second, fixed by a third, verified by a fourth. Four musicians, zero meetings. The protocol produces pipeline chains that no one designed and everyone depends on.

The Trajectory

What Happens Next

The learning store accumulates patterns across every run. Every score makes the next score smarter. This compounds.

The instrument layer is model-agnostic. Claude today. Whatever is best tomorrow. Whatever is best for each specific task within the same score. The orchestration layer outlives any individual model.

24 self-evolution cycles completed autonomously. Each one measurably better than the last. The system improves its own accuracy when fed itself as input.

One person. A laptop. A YAML file. 32 AI agents. 98,802 lines of production code. 11,397 tests passing. mypy strict. ruff clean.

The question is what happens when more people have access to this.