Muse Code Beta Launches for Long-Horizon Software Engineering

Muse Code, a new terminal-based coding agent, has entered beta, aiming to streamline long-horizon software engineering tasks. This tool, powered by the updated Muse Spark 1.2 model, is designed to autonomously plan, execute, and validate intricate, multi-file code modifications across extensive repositories. It leverages persistent sub-agents to enhance problem-solving efficiency and accuracy, minimizing the need for constant human intervention. The underlying architecture of Muse Code incorporates a simple agent loop complemented by asynchronous background agents that remain active throughout a session. This design prevents redundant information gathering and reduces the need for manual steering on complex tasks. An immutable local event log meticulously records every action, ensuring workflows are replay-exact and can safely restart, which is crucial for prolonged development cycles. During rigorous stress testing, Muse Code successfully optimized GPU kernels over thousands of tool calls, sometimes running for up to 24 hours, on Nvidia Hopper GPUs. This resulted in competitive performance improvements for specific deep learning operations. The Muse Spark 1.2 model, co-trained with Muse Code, boasts enhanced coding capabilities, including improved code generation, debugging, and end-to-end developer workflows, and is now available through Muse Code and the Meta Model API.