Meta has launched the beta version of Muse Code for large codebases.

Summary

Meta has launched Muse Code, a beta terminal-based AI coding agent powered by the Muse Spark model. It handles complex tasks in large codebases by spawning parallel sub-agents in isolated working trees.

Why it matters

Meta, long considered behind OpenAI and Anthropic in developer-focused AI tools, is entering the coding agent space with a clear cost advantage. Its parallel sub-agent architecture addresses a key pain point for enterprise developers struggling with massive codebases. This move also signals Meta's broader strategy to expand into the enterprise AI market beyond its core advertising technology.

Highlights

• Core architecture: Distributes tasks to parallel sub-agents in isolated working trees to ensure the working copy is never touched.
• Zuckerberg's demo: Six game features were developed simultaneously without conflicts.
• Alexandr Wang positions the product as a cost-competitive alternative to OpenAI's Codex and Anthropic's Claude Code.

Critical take

The claim of developing 'six features simultaneously without conflicts' is based on a single anecdotal test, not a reproducible benchmark. Furthermore, the article provides no independent performance data or direct comparison metrics against Codex or Claude Code. Since the product is still in beta, the 'competitive' framing remains largely aspirational.