이미지: X — 프론티어랩
Summary
- Meta's official X account, AI at Meta, has unveiled the beta of Muse Code, a coding agent that runs in the terminal.
- Muse Code is powered by the newly announced Muse Spark 1.2 model and is aimed at long-horizon software engineering, the company said.
- It was described as planning, implementing, and validating changes across multiple files in large repositories, using an architecture built around persistent sub-agents.
- 발표 주체
- AI at Meta 공식 X 계정
- 발표 시각
- 2026-08-05 19:25 UTC
- 제품명
- Muse Code (beta)
- 형태
- 터미널에서 동작하는 코딩 에이전트
- 구동 모델
- 신규 모델 Muse Spark 1.2
- 지향 과제
- long-horizon 소프트웨어 엔지니어링
- 핵심 동작
- 대규모 저장소의 복잡한 다중 파일 변경을 계획·구현·검증
- 구조적 특징
- 지속형 서브 에이전트(persistent sub-agents) 활용
A coding agent that runs in the terminal
On August 5, 2026 (UTC), Meta's official X account, AI at Meta, unveiled the beta of Muse Code, a terminal-based coding agent. Meta said the agent was designed for "long-horizon software engineering" (AI at Meta). In other words, the company explained, it is aimed not at single-turn code generation but at development work that requires multiple steps to complete.
Powered by the new Muse Spark 1.2 model
Muse Code runs on Muse Spark 1.2, a new model Meta announced alongside it. Details such as the model's parameter count, training data, and benchmark scores were not disclosed in the post. The scope of release, availability by region, and access conditions have also not been confirmed.
Multi-file changes and persistent sub-agents
Meta said Muse Code plans, implements, and validates complex multi-file changes across large repositories. It also described an architecture featuring persistent sub-agents that divide up the work. However, the original post's text is cut off after a link, so it is unclear from the source exactly what specific problems the sub-agents are meant to solve.
| Item | Confirmed details |
|---|---|
| Product | Muse Code (beta) |
| Runtime environment | Terminal |
| Underlying model | Muse Spark 1.2 (new) |
| Scope of work | Multi-file changes across large repositories |
| Processing stages | Plan → Implement → Validate |
| Architectural feature | Persistent sub-agents |
| Pricing / release scope | Not disclosed |
Other examples of terminal-based coding agents and approaches to long-horizon task handling can be found in METAL LAB's coding agent article collection.
What remains unconfirmed
This announcement is based on a single X post, and no other cross-verifiable reporting has been confirmed yet. Benchmark figures, comparisons with existing coding agents, the beta application process, and supported languages and runtimes were not included in the source post. Detailed specifications are expected to be confirmed when Meta releases follow-up documentation or a technical blog post.
