Introducing Muse Code (beta), a terminal coding agent built for long-horizon sof

Muse Code is Meta's coding agent for the terminal, a program that takes a task, plans changes to a codebase, edits the files and runs commands to check the result. Meta introduced it as a beta on August 5, running on Muse Spark 1.2, a model released the same day that Meta's release article says was trained together with Muse Code. Meta aims it at long software tasks that span many files in large repositories.

Its distinguishing design is a set of helper agents that persist for the whole session. The release article describes a main agent loop alongside specialized background agents that stay active instead of being started for a single task and discarded. They carry out next steps and decide when to report back to the main agent, and Meta says their persistence avoids gathering the same information twice. Meta says this lets Muse Code solve difficult problems faster, more accurately and with less human steering, but neither the announcement post nor the article's text gives a measurement for those claims.

Meta's own long-running test in the same release was a GPU kernel optimization run lasting up to 24 hours, which measured kernel speed against a baseline rather than what the persistent agents contributed. Meta's current documentation, read in November, describes the tool running interactively or unattended in scripts and CI pipelines, with approval prompts and an operating-system sandbox switched on from the first run, and still lists Muse Spark 1.2 as the default model.