Skip to content

AI

Meta's new coding agent hires its own interns to do the work

Muse Code, now in beta, plans, writes and checks code in your terminal. For big jobs it spawns sub-agents that work in parallel, which is management, but make it software.

Cover photo: IHazACatNamedMax / CC BY-SA 4.0

TL;DR

  1. Meta released Muse Code, a terminal coding agent, in beta.
  2. It runs on Muse Spark and installs with a single command.
  3. Big jobs get split across parallel sub-agents in isolated worktrees, says Zuckerberg.

What happened

Meta released Muse Code this week, a terminal-based coding agent now in beta. It is aimed at programmers who need help with complex tasks across large software repositories.

Mark Zuckerberg announced it in a social media post on Wednesday. He described it as able to complete full software engineering tasks across big repos, from planning changes to writing code to validating the results.

Under the hood sits Muse Spark, Meta's previously released coding model. Installation takes a single command, so the barrier to trying it is roughly one line of typing.

The interesting part is how it scales. When a job is big enough, Zuckerberg says, Muse Code fans out to separate sub-agents that work in parallel in isolated worktrees. Your working copy is never touched, he added.

His proof point: in testing, Meta had it build several features for a game at the same time with no collisions. That is a vendor demo, not an independent benchmark, so treat it as a promise rather than a verdict.

Why it matters

Meta is widely seen as a straggler in the AI agent race, and this is a catch-up move. It puts the company in the same lane as OpenAI's Codex and Anthropic's Claude Code, with a pitch of being more competitive and more affordable.

The design choice is the story. One agent that does everything in sequence is slow on a huge codebase. Spawning workers that stay in their own sandboxes is how you avoid agents stepping on each other's edits, or on yours.

Your working copy is never touched, Zuckerberg says. That is the whole trust pitch.

It is still an incremental product, not a breakthrough. The terminal-agent category is crowded, and the real test is whether Muse Code holds up on messy, real-world repositories rather than a game demo.

Meanwhile, the carbon-based engineering staff gets a new coworker who never asks for a code review extension.

What to do with it

  • Try it on a throwaway repo.The install is one command, so run it on something you can afford to break before pointing it at production.
  • Check the isolation claim yourself.Verify that your working copy really stays untouched while sub-agents run, instead of trusting the keynote.
  • Compare it head to head.Give Muse Code, Codex and Claude Code the same multi-file task and see who plans, writes and validates best.
  • Mind the beta label.Keep it away from critical branches until Meta has more than a demo to show.

Keep up, sapiens. The interns are now software, and they work in parallel.