Meta Announces Muse Code Terminal Agent and Muse Spark 1.2 Model

ArticleMeta AI ResearchAnnouncement

Meta AI Research has launched the beta of Muse Code, a terminal coding agent driven by their new Muse Spark 1.2 model. The tool handles complex engineering tasks across large repositories using persistent asynchronous background agents and a crash-resilient runtime design.

Key points

  1. Muse Code is a CLI agent for macOS and Linux designed to plan, implement, and validate code across large repositories.
  2. Persistent async background agents remain active throughout a session to reduce latency, prevent duplicate info gathering, and limit steering needs.
  3. A local event log records every model call, tool execution, and edit, enabling exact replays and crash-safe restarts on long-running tasks.
  4. Built-in skills include /plan for gated blueprints, /grill for stress-testing plans, and /goal for task completion.
  5. Muse Spark 1.2 features scaled-up compute on coding benchmarks, co-trained specifically with Muse Code tool harnesses.
  6. A GPU kernel optimization case study demonstrated iterative improvements over 1,000+ tool calls on NVIDIA Hopper architectures for KDA and MLA kernels.

Why it matters

AI coding tools are shifting from interactive single-prompt assistants to fully agentic systems capable of tackling multi-hour software engineering problems. Meta's combination of persistent subagents and state-retaining event logs directly targets the stability issues common in long-horizon autonomous workflows.

Analysis

Meta's release of Muse Code challenges existing terminal-first coding agents by focusing heavily on long-horizon reliability. Keeping background agents persistent rather than spawning them per task reduces information-gathering latency. Furthermore, co-training the underlying model (Muse Spark 1.2) alongside the agent's harness and tools ensures tighter instruction-following during multi-step tasks, as evidenced by their 24-hour GPU kernel compilation benchmark.

Mentioned links

Found in

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Article:Meta Superintelligence Labs

    Meta Superintelligence Labs Unveils Muse Spark Multimodal Reasoning Model

    Meta has announced Muse Spark, the inaugural model in its new Muse lineup. Designed as a natively multimodal reasoning model, it introduces parallel multi-agent orchestration and significantly improved training efficiency.

    AI & LLMs· Announcement

  • Link:docs.warp.dev

    Warp: Open-Source Terminal as an Agentic Development Environment

    Warp positions its platform as an Agentic Development Environment. The Rust-based terminal combines command-line workflows with AI agents capable of operating interactively inside the app, via a standalone CLI, or asynchronously in the cloud.

    AI & LLMs· Tool

  • Repository:xai-org/grok-build

    Grok Build: Terminal-Based AI Coding Agent by SpaceXAI

    Grok Build is a Rust-based AI coding agent featuring a full-screen, mouse-interactive TUI. It analyzes codebases, modifies files, runs shell commands, and supports interactive use, headless CI scripting, and editor embedding via the Agent Client Protocol.

    26.5kstarsRust

    AI & LLMs· Tool

  • Link:Austin Wang, Lawrence Chen

    cmux: An Open-Source Terminal Built for Coding Agents

    cmux is a native macOS terminal based on Ghostty, designed specifically for multitasking with coding agents. It features vertical tabs, visual notifications, a built-in browser, and a CLI for automated control.

    AI & LLMs· Announcement

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.