Source: VentureBeat — 2026-08-05
Summary
Meta Superintelligence Labs released Muse Code (beta), a terminal-based coding agent comparable to Claude Code and OpenAI's Codex CLI, powered by a new model called Muse Spark 1.2. It supports repo inspection, parallel subagents that each run in their own isolated git worktree, and crash recovery for persistent async background agents. Meta reports a Terminal-Bench 2.1 score of 82.9%, ahead of GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%) but behind Claude Opus 5 at max effort (86.7%) — landing Muse Code mid-pack among frontier coding agents. These figures are Meta's own self-reported evaluation results and have not yet been independently verified on a public leaderboard.
Key Takeaways
- Muse Code (beta) is Meta's first entry into the terminal coding-agent category, positioned directly against Claude Code and Codex CLI.
- It runs on Muse Spark 1.2, a new model from Meta Superintelligence Labs built specifically for this agent.
- Each parallel subagent gets its own isolated git worktree, so multiple agents can edit files simultaneously without overwriting each other's changes — the core mechanism that makes safe parallelism possible.
- Crash recovery lets persistent async background agents resume after a failure instead of losing progress and restarting from scratch.
- Meta's self-reported Terminal-Bench 2.1 score of 82.9% beats GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%), but trails Claude Opus 5 at max effort (86.7%).
- Caveat: all of these benchmark numbers come from Meta's own evaluation, not an independent public leaderboard — treat the ranking as a claim, not a confirmed result.
Reel Script
Hook: Meta just released a coding agent that claims to beat GPT-5.6 and Grok 4.5 head to head on a real coding benchmark. Before you switch your workflow over, you should know Meta graded its own homework on that scorecard — nobody independent has checked the number yet.
Core Concept: The genuinely interesting part isn't the benchmark ranking, it's how Muse Code handles parallel subagents. Modern coding agents increasingly split a big task across multiple subagents working at once — one refactoring a module, another writing tests, another updating docs. The problem is they all normally share one working copy of the repo, so two agents editing files at the same time can literally overwrite each other's work. Muse Code's fix is to give every subagent its own isolated git worktree — a git worktree is a separate working directory that's still linked to the same repository and history, like handing each subagent its own physical copy of the blueprint to mark up instead of five people scribbling on one master copy at the same time. Each subagent does its work in isolation, and the changes get reconciled back into the main branch once it's done, so parallel work doesn't collide. On top of that, Muse Code adds crash recovery for persistent async background agents — meaning a long-running task that dies partway through can pick back up where it left off instead of restarting from zero, which matters once you're running agents unattended for hours.
Hands-On: The benchmark table, per Meta's own numbers on Terminal-Bench 2.1 — a benchmark that scores agents on completing real terminal-based coding tasks end to end: Claude Opus 5 at max effort leads at 86.7%, Muse Code comes in at 82.9%, GPT-5.6 Terra scores 81.8%, and Grok 4.5 scores 81.6%. So by Meta's own scoring, its debut coding agent lands solidly mid-pack — ahead of two frontier competitors, clearly behind the top one. The architecture flow worth picturing: a main agent receives a task, decides to split it across subagents, each subagent spins up its own git worktree off the same repo, works independently, and the results merge back into the primary branch — with crash recovery watching the whole pipeline so a dropped connection doesn't wipe out an hour of agent work.
Takeaway: The worktree-isolation idea is legitimately solid engineering regardless of where the benchmark shakes out — it's a clean solution to a real concurrency problem in agentic coding tools. But an 82.9% that Meta measured itself isn't the same as an 82.9% on a neutral leaderboard, so hold the ranking loosely until someone outside Meta reproduces it. Worth trying Muse Code for the worktree parallelism alone, just don't take the benchmark table as gospel yet.