Muse Code is Meta's beta terminal coding agent, powered by Muse Spark 1.2. Its distinguishing design combines a simple foreground agent loop with persistent asynchronous background agents and an append-only local event log intended to make long-running sessions restart-safe and exactly replayable.
What It Is
Muse Code plans changes, edits code, and validates results across large repositories. Unlike task-scoped subagent designs, its specialized background agents remain active throughout a session, choose their own next steps, and decide when to report back to the main agent. Meta says this persistence reduces repeated discovery and steering on long, multi-step tasks.
The runtime appends every model call, tool run, approval, and edit to a local event log. That log acts as the source of truth for replay and recovery: after a crash, the agent can resume from the recorded state rather than reconstructing the session from a summary.
Bundled workflows include:
/plan: produce an approval-gated implementation plan./grill: challenge the plan before execution./goal: keep working toward an explicit completion objective.
Co-Designed Model and Harness
Muse Code ships with Muse Spark 1.2, a coding-focused update to Meta's Muse Spark model. Meta reports co-training the model with the harness using sampled trajectories and training recipes for goals, compaction, subagents, and tool use. The launch materials also describe author-run kernel-optimization sessions lasting more than 1,000 tool calls and up to 24 hours.
That co-design is strategically important: it treats model behavior, tool affordances, context management, and long-horizon execution as one system instead of assuming a general model will adapt equally well to every harness.
Why Assess
Muse Code is a credible new entrant from a major model provider, but it is still a beta with only vendor-reported evidence:
- No independent production evidence: the public launch shows demos and internal evaluations, not named external deployments or longitudinal reliability data.
- Persistent-agent coordination risk: always-on background agents may reduce latency, but they also increase concurrency, observability, and stale-context risks.
- Replay claims need operational testing: an append-only event log is a strong recovery primitive, but teams should verify idempotency around partially completed tool calls and external side effects.
- Model-harness coupling: co-training may improve performance in Muse Code while making it harder to separate model gains from harness gains or to substitute another provider.
- Availability is early: Muse Code is labeled beta, and Meta's announcement does not yet provide a mature enterprise governance or deployment record.
Assess it on bounded, disposable repositories. Measure crash recovery, approval replay, background-agent coordination, duplicate side effects, and the quality of /grill feedback before considering production work.
Key Characteristics
| Property | Value |
|---|---|
| Interface | Terminal coding agent |
| Provider | Meta |
| Status | Beta |
| Underlying model | Muse Spark 1.2 |
| Coordination | Persistent asynchronous background agents plus a main agent loop |
| Recovery | Append-only local event log; restart-safe replay design |
| Built-in workflows | /plan, /grill, /goal |
| Platforms | macOS and Linux |
| Announcement | Introducing Muse Code and Muse Spark 1.2 |
Sources
- Introducing Muse Code and Muse Spark 1.2 — official beta announcement, architecture, training approach, availability, and author-reported case studies