Meta Releases Muse Code Coding Agent With Muse Spark 1.2

Muse Code combines a primary agent loop with persistent subagents that can remain active throughout a task. Meta says the design reduces repeated information gathering and helps the system coordinate long-running repository work.

The runtime also writes model calls, tool runs, approvals, and edits to an append-only local event log. Meta describes that log as the source of truth for replaying and resuming work after a failure. For engineering teams, that design is more useful than an opaque transcript because it creates a record of which tools ran, which changes were approved, and how the task progressed.

The beta includes workflow controls for planning and review. Meta describes an approval-gated /plan command, a /grill command that stress-tests the plan, and a /goal mode that continues toward a specified result. Those controls can improve observability, but they do not replace human review of proposed changes, tool permissions, tests, secrets, or security-sensitive operations.

Muse Spark 1.2 is Meta's coding-focused update to the Muse Spark family. The company says it increased coding-task training compute, diversified the training environments, and co-trained the model with the Muse Code harness. Meta also says the model was trained for long-horizon tasks using planning, goal conditioning, and context compaction. These are vendor-reported design and performance claims; independent evaluation will be needed to establish reliability on production repositories.

The practical test is therefore not only whether Muse Code can finish a benchmark. Teams evaluating the beta should measure whether its event log, approval gates, validation loop, and recovery behavior make long-running agent work observable and reversible under their own repository policies.