Meta Muse Code

Proceed with caution

Why

The architecture is a direct match for our conductor model — fan-out to sub-agents in isolated git worktrees (one worktree per child from the lead's commit, concurrency cores-2 clamped 2-16) plus an append-only event log giving crash recovery and exact replay is precisely the autonomous-conductor pattern we already run by hand. But adoption evidence is weak in three ways: chart extractions of Meta's own published comparisons put Muse Spark 1.2 second to Claude Opus 5 on all three benchmarks (82.9 vs 86.7, 59.3 vs 65.0, 70.6 vs 79.4), the one third-party-verified predecessor number came in 3.8 points below the claim, and the pricing that makes it attractive is paid for with training rights over the submitted codebase. Worth a measured trial, not a default-implementer swap.

Next

Run a bounded head-to-head on a NON-proprietary repo using the Standard tier only ($1.25/$4.25 per M, no training rights): fixed task set against the current Codex + Claude Code implementers, measuring wall-clock, conflict rate on parallel worktree fan-out, and cost per merged change. Re-evaluate adoption when an independent Terminal-Bench 2.1 verification for Muse Spark 1.2 is published by a benchmark authority.