Meta Launches Muse Code and Muse Spark 1.2 for AI Coding
AI coding agents can write a function quickly. The harder engineering work begins when a task spans an entire repository, requires repeated testing and continues long enough for tools or sessions to fail.
Meta is targeting that problem with Muse Code, a terminal-based coding agent released in beta alongside its new Muse Spark 1.2 model.
Muse Code is designed to plan software changes, write code and validate the result across large repositories. Its distinctive feature is a group of persistent background agents that remain active throughout a session instead of being created separately for every task. Meta says this reduces repeated information gathering and gives the main agent more support during complex, multi-step work.
The runtime also records every model call, tool execution, approval and edit in a local append-only event log. That log acts as the system’s source of truth, allowing Muse Code to resume from the same point after a crash instead of restarting a long task.
Muse Code includes several built-in workflows. The /plan skill creates an approval-gated implementation plan, /grill challenges the plan for weaknesses, and /goal continues working toward a defined outcome. The agent is currently available for macOS and Linux.
Muse Spark 1.2 is built for repository-scale work
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1. Meta says it increased training compute for coding and expanded the range of training environments, targeting improvements in code generation, debugging, codebase understanding and complete developer workflows.
The model was co-trained with Muse Code rather than developed separately and connected later. Training included agent trajectories, tool integration, context compaction and subagent coordination. It was also trained on long-horizon tasks such as generating complete repositories, executing large projects and conducting automated research.
Meta tested the system on external benchmarks including Terminal-Bench 2.1 and DeepSWE 1.1, as well as an internal coding benchmark based on 440 tasks drawn from real Meta pull requests. The company cautions that comparisons with third-party models used different agent products and may not represent each competing model’s best performance.
Muse Spark 1.2 is available through Muse Code and the Meta Model API. The launch places Meta directly in the market for AI coding agents that can move beyond code suggestions and execute longer software-engineering workflows.