openmythos — OpenMythos (kyegomez/OpenMythos)

status: insufficient-public-material-for-agent-framework

OpenMythos (13,396 stars) is a PyTorch implementation of a theoretical Recurrent-Depth Transformer (RDT) architecture, described as "a theoretical reconstruction of the Claude Mythos architecture, built from first principles using the available research literature." It is NOT an AI agent framework, harness, or orchestration system. It is a deep learning model architecture with three stages: Prelude (transformer blocks), a looped Recurrent Block, and a Coda, with switchable attention (MLA or GQA) and sparse Mixture-of-Experts feed-forward. The repo ships pre-configured scales from 1B to 1T parameters, a training script for the 3B model on FineWeb-Edu, and a pip-installable package (open-mythos).

This repo was likely included in batch-23 due to the "AI-as-a-system" / "OS" keyword cluster ("Mythos" + "OpenMythos" + kyegomez's reputation for agent frameworks), but it is a model architecture repository, not a framework for coding agents. Writing a full 11-file report would require inventing content not present in the repo.