AI News

Muse Code: Meta's Coding Agent and Muse Spark 1.2

#Meta#Coding AI#AI Agents#Terminal

What is Meta Muse Code, the coding agent now in beta?

Muse Code is a terminal coding agent that Meta released in beta on August 5, 2026. It works on large repositories: it plans a change, writes the code and validates the result. A coding agent is a program that carries out those steps itself rather than only suggesting code.

Muse Spark 1.2, which Meta called its newest model on August 5, powers the agent. Because Muse Code is a terminal agent, it runs in a shell rather than inside an editor. Meta announced the agent and the model together.

We're excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model. ／ Muse Code takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results. — From the opening paragraph and the install section of the announcement
View official source →

Muse Code release date and install on macOS and Linux

Meta announced Muse Code on August 5, 2026, the same day as Muse Spark 1.2. It installs on macOS or Linux with one line in the terminal:

curl -fsSL https://dev.meta.ai/install.sh | bash

The command downloads Meta's install script with curl and hands it to bash. The install instructions in the announcement name macOS and Linux.

Meta's announcement lists Muse Spark 1.2 as available that day in Muse Code and in the Meta Model API, with expanded global access. The model is therefore offered outside the agent as well.

Install Muse Code on macOS or Linux: curl -fsSL https://dev.meta.ai/install.sh | bash ／ Availability ／ Muse Spark 1.2 is available today in Muse Code and in Meta Model API with expanded global access. — From the install command and the Availability section
View official source →

Meta Muse Code beta: async subagents and bundled skills

Muse Code runs a simple agent loop and adds a set of async background agents that strengthen the main agent. For each task it can coordinate multiple persistent subagents, meaning separate agent instances that the main agent directs. Meta says this solves hard problems faster, more accurately and with less intervention.

The beta also ships with several default skills, which are ready-made commands for common steps. Meta names three, and /grill works on the plan that /plan produces:

Skill What it does
/plan Turns a task into a plan that needs approval before work continues
/grill Stress-tests that plan until it holds up
/goal Works toward successful completion of the objective you specify
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 agent loop, the subagent design and the bundled skills
View official source →

Meta coding agent event log: Muse Code resumes after a crash

Meta says Muse Code can resume where it stopped after a crash, because it keeps a local event log. Every model call, tool run, approval and edit is appended to that log, and Meta calls it the single source of truth for the runtime.

Say the session crashes when a change across a large repository is half finished. An agent with no record of its steps has nothing to resume from. Muse Code has that record, and Meta says it can pick up precisely where it stopped.

Meta describes the result as a runtime that is replay-exact and restart-safe. It adds that this lets Muse Code keep working on long tasks even when something fails.

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
View official source →

Meta Muse Spark 1.2: the coding model behind Muse Code

Muse Spark 1.2 follows Muse Spark 1.1 with a focus on coding. Meta says it improved how the model writes code, debugs hard problems, reads a codebase and handles developer workflows end to end.

Meta co-trained the model with Muse Code so the model performs best, and is easiest to code with, when the two are paired. Meta also trained the model on long-horizon coding tasks, which need many steps before the work is done:

  • whole-repository generation
  • large end-to-end projects
  • auto-research

Meta says the model uses planning to order its work and goal conditioning to hold its direction. It also uses context compaction to keep the knowledge it needs to sustain progress.

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. ／ We co-trained Muse Spark 1.2 with Muse Code to ensure the model exhibits its best performance and coding usability when paired together. ／ 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 model update, co-training with Muse Code and long-horizon training
View official source →

FAQ

What does Muse Code do?
It is a coding agent that runs in the terminal. It targets complex software engineering work spanning large repositories, covering planning the changes, writing the code and validating the results. The model behind it is Muse Spark 1.2, and it was released as a beta on August 5, 2026.
If it crashes mid-task, does the work start over?
Not according to Meta. Muse Code appends every model call, tool run, approval and edit to a local event log, and treats that record as the single source of truth. Meta states this makes the runtime replay-exact and restart-safe, so after a crash the agent can resume precisely where it stopped.
Where can I run it, and on which operating systems?
You install it on macOS or Linux. The method Meta documents is a single line that fetches the install script with curl and pipes it to bash. Meta's announcement lists Muse Spark 1.2 itself as available in Muse Code and also through the Meta Model API.