Meta Releases Muse Code with Muse Spark 1.2 Model

Meta has introduced Muse Code (beta), a terminal coding agent powered by the new Muse Spark 1.2 model. It's designed to handle complex software engineering tasks across large repositories, from planning to validation, and is available for macOS and Linux via a curl install script.
Async Background Agents
Muse Code runs a main agent loop but also spawns persistent async background agents that stay active for the entire session. Instead of being created per task, these agents maintain context and pick up next steps autonomously, communicating when needed. This design cuts down on redundant information gathering and reduces the need for manual steering on multi-step tasks.
Replay-Exact Runtime
A core feature is the local event log, which records every model call, tool run, approval, and edit. Because all actions are logged, the runtime is replay-exact and restart-safe: if the process crashes, it can resume exactly where it stopped. That makes long-running tasks viable without losing progress.
Bundled Skills
Muse Code ships with three default skills:
/plan— turns a task into an approval-gated plan/grill— stress-tests the plan until it holds up/goal— works toward successful completion of the objective
Muse Spark 1.2 Model
The underlying model, Muse Spark 1.2, is a coding-focused update over 1.1. Meta scaled up training compute on coding tasks and expanded training environment diversity. It also maintains strength in general agent tasks. The model was co-trained with Muse Code to ensure best performance in the harness, and it's trained on long-horizon coding tasks like whole-repository generation, large end-to-end projects, and auto-research.
Case Study: GPU Kernel Optimization
In a kernel optimization test, the agent ran for over 1,000 tool calls (up to 24 hours) on NVIDIA Hopper GPUs. It iteratively wrote, compiled, profiled, and improved GPU kernels for KDA and MLA workloads, with the model forbidden from importing third-party kernel libraries. Muse Spark 1.2 produced two-kernel Triton pipelines, achieving significant performance gains over baselines by applying optimizations like re-centering gated cumulative decay for KDA and reusing the shared KV latent for MLA.
Availability
Muse Spark 1.2 is available today in Muse Code and via Meta's platform. To install Muse Code, run:
curl -fsSL https://dev.meta.ai/install.sh | bash📖 Read the full source: HN AI Agents
👀 See Also

OpenMontage: Open-Source Agentic Video Production System for AI Coding Assistants
OpenMontage is an open-source video production system that transforms AI coding assistants like Claude Code into full production studios. It handles research, scene planning, script writing, voice narration, music selection, subtitle generation, and validation through automated pipelines.

AgentPVP: An agent-first competitive LLM arena with ELO, rivalries, and prompt-injection sandbox
AgentPVP lets LLM agents register, play 5 board games over JSON APIs, maintain per-game ELO, write rivalry files, and flame each other in a global lounge. HTML is optional — the API is the site.

Homebutler: OpenClaw Skill for Homelab Management via Telegram
Homebutler is a single Go binary (~13MB, zero dependencies) that works as an OpenClaw skill to manage homelabs from Telegram chat. It monitors servers, restarts Docker containers, wakes machines, scans networks, and alerts on resource spikes without SSH sessions or dashboard logins.

ProofShot CLI Gives AI Coding Agents Browser Verification Capabilities
ProofShot is an open-source CLI tool that lets AI coding agents verify UI features by recording browser sessions, capturing screenshots, and collecting console errors. It works with any agent that can run shell commands and generates self-contained HTML reports for human review.