what is writing-driven-autoresearch fr?

This repository is the complete record of a multi-agent system that wrote an academic research paper from start to finish in about three hours, winning first place and 10,000 dollars in OpenAI credits at a hackathon judged by eleven expert reviewers. Almost every file in the repository, including the experiment code, the paper itself, and a running log of decisions, was written by AI agents rather than a human. The only file a person wrote is the README. Three separate AI agents split the work with strict boundaries. One agent owns the overall clock and makes decisions but runs no code. A second agent owns the experiments, running models on GPUs and saving results as JSON files, but never touches the paper. A third agent owns the paper itself and never touches a GPU. The agents never talk to each other in chat. Instead, everything that crosses from one agent to another is written to a shared folder of files, including a status file and a decisions log, so nothing important depends on a message being remembered correctly. The core method is to write a complete, submittable draft of the paper before any experiment has produced real results, using two special markers in the paper's text: one for a number that is still just a guess, and one for a number that is backed by an actual measured result saved on disk. A script checks every measured number against its backing file, and another script fails the process if any guessed number is still present when the paper is due. This forces every claim in the final paper to be either deleted or backed by a real result, while still letting the team keep a paper that could be submitted at any moment as the plan changed underneath it. The repository includes the three agent prompts, the paper specification and style guide, a detailed project plan with a planned response for every possible experimental outcome, the full decision log, and the verification scripts. The authors present it as something to copy and adapt for running the same kind of agent driven research process on a different project and topic.