Meta's Muse Code Tackles 24-Hour Coding Jobs Where Other AI Agents Quit
Meta launches Muse Code, a terminal coding agent with persistent sub-agents and a co-trained model that can grind GPU kernels for 24 hours straight
- Meta launched Muse Code (beta), a terminal coding agent for macOS and Linux, installable with one curl command.
- Powered by Muse Spark 1.2, a co-trained model scoring 82.9% on Terminal-Bench 2.1 and 59% on DeepSWE 1.1, beating Grok Build 4.5 and Gemini 3.6 Flash.
- Persistent async background agents stay active throughout a session, with a crash-safe append-only event log that lets the agent resume after failures.
- Stress-tested over 1,000+ tool calls across 24 hours, autonomously optimizing GPU kernels on NVIDIA Hopper hardware with substantial gains over Triton baselines.
- Pricing via Meta Model API: $1.25/$4.25 per million input/output tokens (standard) or $0.10/$0.20 (contributor tier, data shared with Meta).
- No GUI or IDE integration yet; purely terminal-based, unlike Claude Code or OpenAI Codex which have visual interfaces.
Meta's Superintelligence Labs just dropped Muse Code (beta), a terminal coding agent built for the kind of software engineering work that breaks most AI tools: large repos, multi-file changes, and tasks that take hours rather than seconds. It ships alongside Muse Spark 1.2, a coding-focused model update that was co-trained specifically to run inside Muse Code's agent harness.
A new entrant in a crowded arena
The terminal coding agent space is already packed. Unlike ChatGPT Codex and Claude Code, there is currently no dedicated app interface for Muse Code -- it lives entirely in your terminal. That is a deliberate design choice. Muse Code is a terminal agent that installs with a single command and takes on complete software engineering tasks, including planning changes, writing code, and validating results. The install command is literally one line:
curl -fsSL https://dev.meta.ai/install.sh | bashThe release follows Meta's launch of Muse Spark 1.1, which upgraded the model's agentic and multimodal abilities and introduced Meta's paid API service for developers. With Muse Code, Meta is now competing directly with Claude Code and OpenAI Codex on the autonomous, long-horizon coding agent front.
The architecture: agents that don't quit
The core insight behind Muse Code is that most coding agents fail on long tasks not because they're dumb, but because they're stateless. Every hard problem requires re-gathering context, re-planning, and re-steering. Muse Code attacks this with two architectural ideas.
First, persistent background agents. Muse Code uses persistent background agents, repository-scale execution, and built-in verification. Muse Code operates specialized background agents that remain active throughout a user's session , rather than being spun up and torn down for each subtask. They carry out next steps autonomously and choose when to report back to the main agent, cutting down on redundant context-gathering and the need for human steering.
Second, a crash-proof event log. Muse Code uses a local event log in which every model call, tool run, approval, and edit is appended. This single source of truth makes the runtime replay-exact and restart-safe: after a crash, the agent can resume precisely where it stopped. That ability lets Muse Code take on long-running tasks without being derailed by failures. For tasks that run for hours, this is not a nice-to-have -- it is the difference between finishing and starting over.