Meta Launched Muse Code
Meta AI Research ● Covered by 6 sources
Meta just launched Muse Code, a terminal coding agent running on its new Muse Spark 1.2 model. It can juggle background subagents and even survive crashes mid-task.
Based on reporting by Meta AI Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta has pushed out Muse Code in beta, a command-line coding agent built to tackle the kind of software engineering work that spans entire repositories rather than single files. It plans changes, writes the code, and checks its own results, and it's available now for macOS and Linux users willing to try something still labeled beta.
What sets Muse Code apart is how it handles the grunt work. Instead of spinning up a fresh helper for every subtask, it keeps a set of background agents alive for the whole session. These agents work independently and decide on their own when to report back to the main process, which cuts down on repeated information-gathering and reduces how often a human has to step in on gnarly, multi-step jobs. There's also a local event log recording every model call, tool run, approval, and edit. That log makes the whole system replay-exact and restart-safe, meaning if it crashes, it picks up exactly where it left off instead of starting over.
The tool ships with a few built-in skills worth noting: /plan turns a request into a plan that needs approval before it runs, /grill pressure-tests that plan, and /goal just drives toward finishing the stated objective. Meta's own example has a user dropping in a walkthrough video of a house as an mp4, and the agent turning that into a full vacation-rental booking page.
Under the hood is Muse Spark 1.2, a coding-focused refresh of Muse Spark 1.1. Meta says it poured more training compute specifically into coding tasks and broadened the variety of training environments, while keeping the model's general-agent abilities intact. The two products were trained together, with Muse Code's own toolset folded into the training data so the model and the harness fit each other well. Spark 1.2 was also drilled on long-horizon work — full-repo generation, big end-to-end builds, that sort of thing — using planning, goal-conditioning, and context compaction to keep from losing the thread over long sessions. Meta even used the earlier 1.1 model to generate tricky coding challenges and grade candidate solutions, building a training set aimed at making 1.2 better at following complicated instructions.
As a stress test, Meta had the model optimize GPU kernels over more than a thousand tool calls, running as long as 24 hours, on KDA and MLA kernels for NVIDIA Hopper hardware. The models weren't allowed to just import existing kernel libraries; they had to write the optimizations themselves in Triton. Meta reports the agent kept finding real improvements over the baseline implementation. Muse Spark 1.2 is rolling out today inside Muse Code and through Meta's Model API, with wider global access than before.
My take — AI-written commentary, not fact-checked reporting
The persistent-background-agent design is the genuinely interesting bit here, not the marketing copy about vacation rental pages — plenty of coding assistants can write a function, far fewer can survive a crash and pick up exactly where they left off. Still, a 24-hour kernel-optimization demo on curated benchmarks tells people little about whether this thing behaves on a messy, real-world codebase full of undocumented assumptions. Meta calling this a step toward the frontier, while pointedly noting bigger models are coming, reads like a company setting expectations for a sequel before anyone's finished reviewing the original.
Read more about this at: Meta AI Research