Meta ships Muse Code, a coding agent that keeps working after it crashes
Meta released Muse Code in beta for macOS and Linux, powered by a new Muse Spark 1.2 model. Its two interesting ideas are background agents that stay alive for a whole session and a local event log that lets an interrupted job resume exactly where it stopped.
Meta released Muse Code in beta on Wednesday, a coding agent that runs in your terminal rather than in a chat window. Mark Zuckerberg announced it himself. It runs on a new model called Muse Spark 1.2, and it puts Meta directly against Anthropic’s Claude Code and OpenAI’s Codex in what has quietly become the most competitive corner of the industry.
The pitch is complete software tasks across large codebases: the agent plans a change, writes the code, then checks whether the result actually works. It installs on macOS and Linux with a single shell command, and Muse Spark 1.2 is also available through the Meta Model API. There are three built in skills, which are just named routines you can call: /plan turns a request into a plan you have to approve, /grill attacks that plan to see if it survives, and /goal keeps working until the objective is met.
Two design choices are worth understanding even if you never write code. First, the background agents stay alive for the whole session instead of being created fresh for each task, so they do not have to re-read the same files over and over. Second, every model call, tool run, approval and edit gets appended to a local log file, which Meta describes as replay exact and restart safe. In plain terms: if the thing crashes four hours into a job, it picks up precisely where it stopped rather than starting again. Meta showed a case study where the system made more than 1,000 tool calls over runs of up to 24 hours while optimising GPU kernels.
What is actually going on here
The interesting claim is not the benchmark score, it is that Meta trained the model together with the tool that runs it. Most labs build a model and then wrap a harness around it, meaning the scaffolding that lets a model use a terminal, read files and run tests. Meta co-trained Muse Spark 1.2 with Muse Code, so the model has practised inside the exact environment it ships in. Meta is upfront that its harness may not be tuned for competitors’ models, which is a polite way of saying its comparison numbers against Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, Grok 4.5 and Kimi K3 should be read with care.
Worth knowing before you get excited: this is beta software, there are no downloadable weights, so you are depending on Meta’s servers, and installing anything with a curl | bash command means trusting the publisher completely.
What this means for you: if you do not code, nothing today, but the crash resumption idea is a preview of where all AI assistants are going, from long research jobs to overnight document work. If you do code, this is now a genuine fourth option worth a weekend test, especially for long refactors where the agent used to lose the thread. Keep it in a sandbox with review gates, the same as any agent.
Sources
Source: https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
OpenAI ships three education plugins, and quietly admits most students barely use the tools they have
New plugins for K-12 teachers, college educators and college students land in ChatGPT Edu and ChatGPT for Teachers. Buried in the announcement: even advanced student users tap ChatGPT's capabilities 90 to 99 percent less than power users.
