Meta has launched the beta version of Muse Code, a terminal coding agent powered by Muse Spark 1.2. The release marks the company’s latest step in AI developer tooling, with Muse Code designed to work across large repositories by planning changes, writing code, and validating results.

Muse Code is available on macOS and Linux and can be installed through a single command. Meta said the agent can coordinate multiple persistent subagents for each task, helping it address difficult software engineering problems with less manual intervention.

Persistent Agents and Restart-Safe Runtime

Muse Code uses a simple agent loop supported by async background agents. These background agents remain active throughout a session rather than being created for individual tasks, allowing the agents to avoid repeated information gathering, carry out subsequent steps, and decide when to report back to the main agent.

The agent also uses a local event log that records every model call, tool run, approval, and edit. Meta said this creates a single source of truth that makes the runtime replay-exact and restart-safe, allowing Muse Code to resume from the point where it stopped after a crash.

Muse Code ships with default skills for software development workflows. The /plan command turns a task into an approval-gated plan, /grill stress-tests that plan, and /goal works toward completing a specified objective.

Muse Spark 1.2 Powers the Agent

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding and end-to-end developer workflows. Meta said it scaled up training compute on coding tasks and expanded training environment diversity while maintaining the model’s strength in general agent tasks.

Meta also co-trained Muse Spark 1.2 with Muse Code to improve performance and coding usability when the model and agent are used together. The company said the training included harness trajectories, recipe optimizations for goals, compaction and subagents, and integration of the Muse Code toolset.

Why It Matters

Meta said Muse Spark 1.2 was trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research. The model uses planning, goal conditioning, and context compaction to maintain progress across extended software development work.

In a kernel optimization case study, Meta tested the model over more than 1,000 tool calls across up to 24 hours. Using Muse Code’s agentic coding environment, the model wrote, compiled, profiled, and improved GPU kernel performance relative to a baseline implementation of KDA and MLA kernels on NVIDIA Hopper GPUs.

Muse Spark 1.2 is available in Muse Code and through the Meta Model API, with access expanding globally. Meta said it also plans new harness features and more powerful models.