Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by Muse Spark 1.2

Meta AI has officially introduced Muse Code (Beta), a terminal coding agent accessible via a curl installation script on macOS and Linux systems. It is driven by the newly unveiled Muse Spark 1.2 model, which is now available within Muse Code and the Meta Model API alongside expanded global availability. However, the release notes mention no downloadable weights, meaning developers should treat this integration as a hosted dependency.

Muse Code operates using a straightforward agent loop coupled with a set of async background agents. These specialized background entities remain active throughout an entire session rather than spawning individually for each subtask, which Meta explains prevents redundant information gathering. They execute subsequent steps and decide autonomously when to report back to the main agent, reducing latency and lowering the steering overhead required for complex multi-step tasks.

To safeguard against unexpected failures, Muse Code utilizes a local event log. Every model call, tool execution, user approval, and code edit is appended sequentially to this log. Meta refers to this single source of truth as replay-exact and restart-safe. Following a system crash, the coding agent can pick up precisely where it left off, allowing long-running development tasks to survive unexpected interruptions seamlessly.

Three default skills come pre-installed with the agent:

• /plan transforms tasks into structured, approval-gated execution plans.
• /grill rigorously stress-tests the established plan until it is solid.
• /goal drives progression toward the successful completion of the stated objective.

Stock photo for illustration only, not from the actual event

The underlying model, Muse Spark 1.2, represents a code-focused iteration succeeding Muse Spark 1.1. Meta reports measurable performance gains in code generation, complex debugging, repository comprehension, and end-to-end developer workflows. The research team significantly scaled up training compute dedicated to coding tasks and broadened the diversity of testing environments while retaining the model's proficiency in general agent tasks.

Meta's methodology report provides granular evaluation metrics. Terminal-Bench 2.1 evaluates all 89 tasks using pass@1 over five attempts, while DeepSWE v1.1 covers 113 tasks across 91 repositories and five languages. Additionally, Meta Internal Coding Bench incorporates 440 tasks derived directly from real internal pull requests, with all executions isolated inside Daytona cloud sandboxes.

Source: MarkTechPost