What Muse Code is
Muse Code is a terminal coding agent Meta has released as a beta. It is aimed not at one-off code generation but at complex engineering work spanning large repositories — planning the changes, writing the code, and validating what came out.
Installation is a single command on macOS or Linux. The skeleton is a simple agent loop — take an instruction, think, use a tool, look at the result, think again — extended with asynchronous background agents. Because multiple subagents can work in parallel on one task, Meta describes it as solving difficult problems faster, more accurately and with less intervention.
Three skills ship with it. /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal works toward completing the objective you name. The fact that an approval gate is built into the planning stage from the start says a lot about the character of this trio.
We're excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model. ／ Install Muse Code on macOS or Linux: curl -fsSL https://dev.meta.ai/install.sh | bash ／ Muse Code takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results. ／ Async Background Agents ／ Muse Code operates with a simple agent loop plus a set of async background agents to enhance the main agent's capability. ／ It can coordinate multiple persistent subagents for each task, solving difficult problems faster, more accurately, and with less intervention. ／ Bundled Skills ／ Muse Code ships with several default skills. ／ /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal works toward successful completion of the specified objective. — From the passages on the release form and install command, the scope of work, the async subagent design, and the bundled skills
The event log design that survives a crash
With a long-running agent, the practical problem is what happens when it fails partway: you start again from the beginning. Muse Code answers that through how it stores the execution record.
Every model call, tool run, approval and edit is appended to a local event log. Treating that record as the single source of truth makes the run exactly reproducible, and Meta writes that after a crash the agent resumes precisely where it stopped. That property is what keeps long tasks from being derailed by failures.
This is a different axis from narrowing an agent's permissions behind approvals. That approach controls how far the agent may reach; this one is about not losing the work in progress.
Runtime Design ／ 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. — From the Runtime Design section, on what the event log records, reproducibility and restart safety, and the effect on long-running tasks
How Muse Spark 1.2 was built
The other half of the announcement is Muse Spark 1.2, a coding-focused update to Muse Spark 1.1 with improvements in code generation, complex debugging, codebase understanding and end-to-end developer workflows.
Two things stand out in how it was trained. The first is co-training with Muse Code: the model was trained so that the pair performs at its best when used together. That is a different premise from building an agent and a model separately and combining them afterwards.
The second is training built around long horizons. It was trained on whole-repository generation, large end-to-end projects and auto-research, using planning to sequence the work, goal conditioning to keep the direction from drifting, and context compaction to retain the knowledge needed to keep going.
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. ／ Co-Training With Muse Code ／ We co-trained Muse Spark 1.2 with Muse Code to ensure the model exhibits its best performance and coding usability when paired together. ／ Long-Horizon ／ Muse Spark 1.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research. ／ It leverages planning to sequence work, goal conditioning to maintain direction, and context compaction to retain the knowledge needed to sustain progress. — From the passages on the position of Muse Spark 1.2, the aim of co-training with Muse Code, and training for long-horizon work
How to read Muse Code
What characterises Muse Code is less the intelligence of the model than the construction of the harness around it. Resumability through the event log, the approval-gated /plan, the /grill that beats on the plan — each of these exists so that a long unattended run does not break.
If you are evaluating it, the recovery cost when a run fails halfway through your repository will tell you more than the benchmark numbers. Making it easy to inspect diffs locally also speeds up checking whatever the agent wrote.
