Meta releases Muse Code beta terminal agent for large repositories
Meta has released Muse Code, a beta terminal coding agent intended to handle complete software-engineering tasks across large code repositories. Mark Zuckerberg said the tool can plan changes, write code and validate results, positioning it for programmers tackling complex work rather than isolated code prompts.
The agent is installed with a single command and is powered by Muse Spark, Meta's previously released coding model. Meta says Muse Code can distribute work when a task becomes large enough, using multiple agents at the same time.
Parallel agents in isolated worktrees
Zuckerberg described a design in which a large job fans out to separate sub-agents working in parallel in isolated worktrees. The arrangement is intended to leave a developer's working copy untouched while the sub-agents perform their assigned work.
In Meta's testing, the agent built six game features simultaneously without collisions. That example illustrates the product's focus on coordinating several pieces of work inside one repository, alongside its stated abilities to plan, implement and validate changes.
The product builds on Muse Spark advances in agentic coding as Meta develops its Muse Spark model for agentic coding workloads. Muse Code applies that model in a terminal agent aimed specifically at engineering tasks spanning large repositories.
Meta seeks a place in AI coding tools
The launch places Meta alongside AI coding offerings from OpenAI and Anthropic, including Codex and Claude Code. Alexandr Wang, Meta's AI chief and leader of Meta Superintelligence Labs, told the Wall Street Journal that the company sees Muse Code as a strong option for many workflows and use cases, particularly on cost.
Meta has been increasing its AI investment as it expands beyond its core use of AI in advertising. In June, the company entered the enterprise AI market with an agent intended for customer service and support.
Business implication
For software teams, Muse Code is a beta option to evaluate on large repository tasks that can be divided into parallel workstreams. Teams can examine whether its isolated-worktree approach, planning, code generation and validation fit their existing development controls before relying on it for broader engineering work.