Meta has introduced Muse Code in beta, a terminal-based coding agent powered by its new Muse Spark 1.2 model. The tool is designed for extended software-engineering work across large repositories, combining a primary agent loop with background agents that persist for the duration of a session.
Unlike workers created for a single subtask, the background agents remain available and decide when to report findings to the main agent. Meta says that continuity can reduce repeated investigation and the amount of user steering required on long, multi-step changes. The system can use several of these agents to support planning, implementation and verification.
Muse Code records every model request, tool execution, approval and file edit in a local event log. Meta describes that log as the source of truth for the runtime, enabling an interrupted session to replay and resume from the point at which it stopped. That design targets coding jobs that may run long enough for process failures to become a practical concern.
The beta includes commands for structured workflows. `/plan` creates a plan that requires approval, `/grill` challenges the proposal, and `/goal` keeps the agent working toward a specified result. Meta also demonstrated multimodal input by giving the tool a video walkthrough of a home and asking it to produce a marketing and booking page.
Muse Spark 1.2 is presented as a coding-oriented revision of Muse Spark 1.1, with more training compute devoted to programming and a wider variety of training environments. Meta says the model and coding harness were co-trained, including trajectories for goals, context compaction and subagent use. Training tasks covered full-repository generation, larger end-to-end projects and automated research.
One reported evaluation asked the model to optimize GPU kernels over more than 1,000 tool calls and for as long as 24 hours. In the test environment, it wrote, compiled and profiled Triton implementations for KDA and MLA workloads on Nvidia Hopper hardware. Third-party kernel libraries were prohibited, requiring the system to implement the algorithms rather than wrap existing optimized code. Meta reports improvements over the supplied baselines, though the supplied announcement does not provide independent validation.
Muse Spark 1.2 is available through Muse Code and the Meta Model API, with Meta describing access as globally expanded. The release joins a growing category of coding agents built around persistent state, tool use and verification rather than isolated code completion. Its most consequential claims concern endurance and coordination; how those features perform on outside repositories will become clearer as beta users test them under varied constraints.
