Meta Introduces Muse Code and Muse Spark 1.2 Coding Agent
Meta has launched the beta version of Muse Code, its first terminal-based coding agent, alongside Muse Spark 1.2, an upgraded model optimized for complex debugging and long-horizon coding tasks. These tools were co-trained to maximize compatibility and performance in end-to-end developer workflows. This release marks Meta's entry into the competitive landscape of AI coding agents, directly challenging tools like Anthropic's Claude Code. It highlights the industry's shift toward agentic workflows where AI can autonomously execute long-sequence tool calls to solve complex software engineering tasks. During stress testing, Muse Code demonstrated the ability to iteratively optimize GPU kernels over more than 1,000 tool calls spanning up to 24 hours on Nvidia Hopper GPUs. The training process utilized rejection-sampled harness trajectories and recipe optimizations for subagents.
## BACKGROUND
Coding agents are AI systems designed to autonomously write, test, and debug code by interacting with developer environments and external tools. Agentic tool calling refers to an AI's ability to independently choose and execute external functions or APIs to accomplish a goal, which is crucial for handling complex, multi-step software development tasks.