Theo Browne examines Meta's Muse Code terminal agent and Muse Spark 1.2 model. The system keeps background agents active across a session, can split work across isolated worktrees and records model calls, tool runs and edits in a local event log for recovery and auditing.
The model's main advantages are speed and price. Theo Browne sees throughput far above several frontier alternatives, while the contributor tier makes large repository audits exceptionally inexpensive in exchange for allowing Meta to use the submitted data. The standard tier costs substantially more but remains comparatively affordable.
Practical tests reveal an uneven capability profile. Muse quickly audits a repository, generates clear review pages and divides investigations across subagents, yet it also hallucinates the wrong product, struggles with rate limits and fails to complete a complex integration correctly. Independent model reviews rate its implementation plan below a stronger Fable plan.
Theo Browne concludes that Muse is most useful for extracting signals from large amounts of code, classifying pull requests, summarizing logs and generating reports. He does not trust it for long-running implementation work, where a slower but more reliable model can reduce the cost of reviewing and repairing failed changes.
Watch on YouTube
