Meta's Muse Code Runs Six Coding Agents at Once, Each in Its Own Git Worktree

Meta launched Muse Code in beta on August 5, 2026, a terminal-based coding agent built on the company's own Muse Spark 1.2 model. The distinguishing feature is how it handles parallel work: instead of one agent working through tasks sequentially, Muse Code fans out into multiple sub-agents, each operating in its own isolated git worktree, specifically to avoid the merge collisions that happen when several agents edit the same large repository at once.

Meta reportedly demoed the tool building six different game features simultaneously with zero merge collisions, using the parallel worktree structure to keep each sub-agent's changes isolated until they are ready to merge back in. That is a real answer to a problem anyone who has run multiple coding agents against one repo has hit: two agents touching overlapping files and producing a conflict neither one can resolve on its own.

Meta is positioning Muse Code as a lower-cost alternative to Codex and Claude Code, rather than claiming to outperform them outright. For teams already running multiple agent sessions against a single codebase and hitting merge conflicts as a result, the worktree isolation approach is worth watching regardless of which underlying model ends up ahead, since it is solving a coordination problem that exists independent of model quality.