Meta AI has released Muse Code, a beta terminal coding agent for macOS and Linux, powered by the new Muse Spark 1.2 model. The system targets complex software engineering across large repositories by planning changes, writing code, and validating results through a persistent agent loop.
- Muse Code utilizes async background agents that remain active throughout a session to reduce latency and avoid redundant information gathering.
- An append-only local event log records all actions, making the runtime replay-exact and restart-safe after crashes.
- Muse Spark 1.2 was co-trained with the Muse Code harness and focused on long-horizon, repository-scale tasks.
- Evaluation includes Terminal-Bench 2.1, DeepSWE v1.1, and Meta Internal Coding Bench, with comparisons to other major models.
- A case study demonstrated iterative GPU kernel optimization over 1,000+ tool calls lasting up to 24 hours on NVIDIA Hopper GPUs.
The release positions Muse Code as a tool for teams running agents in sandboxes, offering capabilities for repository-scale refactors, long-running bug triage, and test generation.