Meta released Muse Code on Wednesday, a terminal-based coding agent built for engineers working inside repositories large enough to break most AI assistants. The tool entered beta this week and installs with a single command.
Chief executive Mark Zuckerberg described the agent in a social media post as capable of completing whole software engineering tasks across large repos. That framing covers three phases Meta says the agent handles on its own: planning the changes, writing the code, and validating the results afterward. It is a wider remit than autocomplete, and a direct answer to the complaint that AI assistants lose the thread once a task spans more than a handful of files.
Parallel sub-agents in isolated worktrees
The architectural detail Meta is leaning on hardest is how Muse Code splits work. Rather than grinding through a large task in one long session, the agent spawns additional sub-agents when the job justifies it. Zuckerberg said that when a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees.
Worktree isolation addresses the failure mode that plagues agents on large repositories. Multiple agents editing the same checkout produce conflicting writes and half-applied changes. Giving each sub-agent its own working copy lets Meta run them concurrently without results colliding, and it means one sub-agent going off the rails does not corrupt what the others produced.
Muse Code runs on Muse Spark, the coding model Meta shipped earlier and has since updated to version 1.2. Keeping the agent tied to an in-house model gives Meta direct control over what every token costs, which turns out to be central to the pitch.
Competing on price, not just capability
The launch drops Meta into a category OpenAI and Anthropic have largely defined. OpenAI ships Codex, Anthropic ships Claude Code, and both have spent the past year converting developers to terminal-native agents that sit beside a repository rather than inside an editor.
Meta's differentiation is economic. Alexandr Wang, the company's chief AI officer, argued that for a lot of workflows and a lot of use cases the tool can be an incredibly good option, especially from a cost perspective. That is a narrower claim than beating rivals on raw capability, and a deliberate one.
Agentic coding burns tokens at a rate that has made budgeting genuinely difficult for engineering organizations. A single long-running task that plans, edits and re-validates across dozens of files can consume more than a developer's entire monthly allowance, and some teams have throttled agent use rather than absorb unpredictable bills. Undercutting on price is a credible way in.
From publishing models to shipping tools
Muse Code marks how far Meta has moved from releasing model weights toward selling finished developer products built on top of them. The company spent years positioning itself as the open alternative in AI infrastructure. Shipping a packaged agent with its own runtime behaviour and orchestration layer is a different business.
The open question is whether the parallel sub-agent design holds up outside Meta's own monorepo, which is unusually large and unusually well instrumented. Fan-out helps only when a task decomposes cleanly, and messy legacy codebases rarely cooperate. Beta users will settle that quickly, and Meta will need reviewable output rather than volume to keep them.
