On August 5, Meta released the public beta of Muse Code, its first coding agent, available on macOS and Linux. It has a single entry point — the terminal, with no desktop app or editor plugin — installable via a one-line curl command, and is aimed at "complete software engineering tasks across large repositories": planning changes, writing code, and verifying results.
The Model Behind It
Muse Code runs on Muse Spark 1.2, the latest iteration of Meta's post-Llama flagship architecture. Muse Spark debuted in April, version 1.1 (July) added agentic and multimodal capabilities plus a paid developer API, and 1.2 is a fine-tuned refresh specifically trained for debugging, repository understanding, and long-horizon development workflows.
Notably, the model and the agent were designed together — Meta optimized the planning, execution, testing, and verification loop jointly, rather than wrapping a generic model in a shell.
Key Design: Parallel Sub-Agents + Isolated Worktrees
Most coding agents process tasks serially, which becomes a bottleneck on large repositories. Muse Code fans out large tasks to multiple parallel sub-agents, each running in its own isolated git worktree — a second working directory attached to the same repository with its own checkout and index. Sub-agents cannot overwrite each other's files, contend for the index, or disturb uncommitted changes in the developer's working copy. All work is merged back into the main branch afterward.
In Meta's internal testing, six game features were developed simultaneously with zero collisions. Any team can rig this up with git worktree add and a shell loop, but making it default behavior is the real advance.
Second Underrated Design: Local Event Logs
Muse Code logs every step locally: model calls, tool actions, approvals granted, and edits made. The advertised use is crash recovery — a long task that dies after 35 minutes doesn't lose those 35 minutes. But the better use is review: a diff tells you what changed; the log tells you why the agent believed that was the task, under which approval it acted, and what it read before editing. Read the log before merging anything you'd be embarrassed to ship.
Pricing: Read the Page Like a Contract
- Standard tier: $1.25 per million input tokens, $4.25 per million output tokens, $0.15 per million cached input tokens; 3,000 requests/minute and 4M token/minute limits; prompts and completions explicitly not used for training.
- Contribution tier: roughly one-tenth the standard price, but you authorize Meta to train on everything you send — code and prompts.
- Codex (OpenAI) — cloud-based agent, running in enterprise customers for months.
- Claude Code (Anthropic) — terminal agent, the default for companies betting on Anthropic (recently backed by up to $5B in AMD compute commitments).
- Muse Code (Meta) — parallel worktrees + cost, but with no public benchmarks yet (no SWE-bench Verified, long-context recall, or tool-reliability numbers); benchmarks, pricing, and beta status are all still open questions.
The cheap tier isn't free; it's paid in a different currency.
Whose Work Is It Taking?
Alexandr Wang, head of Meta Superintelligence Labs, told The Wall Street Journal the cost-capability balance makes it "a great option for a lot of workflows, especially from a cost standpoint." The competitive landscape:
Verdict
The detail worth remembering isn't "yet another coding agent" — it's the parallel-worktree engineering choice, turning "do parallel agents actually work?" from a slogan into default behavior. Meta's playbook mirrors Llama's open-source strategy: make the model free or cheap first, then compete on the bill. For buyers, whichever vendor wins, the pressure will force OpenAI and Anthropic to cut prices or improve multi-agent capabilities — which is a good thing in itself.
Sources
1. Meta official announcement / Mark Zuckerberg on X, 8/5/2026 2. TechCrunch, 8/5/2026 3. CNBC, 8/5/2026 4. VentureBeat, 8/5/2026 5. peoplearegeek.com technical analysis