Models
Meta's Muse Code Bets Parallel Agents Can Make Large Repositories Manageable
Meta has launched Muse Code in beta, a terminal agent that delegates work across isolated worktrees, putting cost, coordination and verification at the center of its challenge to established coding assistants.
By Patrick T ยท
Meta has released Muse Code in beta, a terminal coding agent aimed at larger software projects where one assistant working sequentially can become the bottleneck. The company says the product, powered by its Muse Spark model, can plan changes, write code and validate results across a repository.
Its core design is parallelism. Meta says bigger jobs can fan out to sub-agents in isolated worktrees so that the developer's main working copy is not touched. That is a sensible response to a practical problem: an agent that understands a broad code base may still create collisions if several tasks edit shared assumptions at once.
Isolation reduces some risk, but it does not make a multi-agent workflow correct by default. Teams still need tests, review rules and a clear account of which agent changed what. The more work is delegated, the more valuable a clean trail becomes when a regression appears after a merge.
Meta is entering a crowded market led by products such as OpenAI Codex and Anthropic Claude Code. Its argument appears to rest partly on cost and partly on the ability to divide a large task without turning one developer session into a queue of long-running requests.
The winner will not simply be the agent that writes the most code. It will be the one that helps a team understand the consequences of that code at repository scale, where the expensive failures are rarely confined to a single file.
Topics: Meta, Muse Code, developer tools