Meta ships Muse Code, a terminal agent that resumes exactly where a crash stopped it
Meta Superintelligence Labs released Muse Code in beta for macOS and Linux, a terminal coding agent whose local event log appends every model call, tool run, and edit so the session restarts exactly where it stopped. Restart-safety is what the release is selling, and Meta's kernel benchmark backs it — the agent optimized KDA and MLA Triton kernels for NVIDIA Hopper across 1,000-plus tool calls and up to 24 hours. Muse Spark 1.2, co-trained with the harness and bundled with /plan, /grill, and /goal skills, ships today in the Meta Model API with expanded global access, moving day-long unattended coding runs from research demo to shipped product.
Source: research.meta.ai ↗
This single source of truth makes the runtime replay-exact and restart-safe: after a crash, the agent can resume precisely where it stopped.
Why this matters
- → Unattended coding runs now survive crashes, enabling 24-hour tasks without manual restart.
- → Iterative optimization over 1,000+ tool calls proves agentic reasoning scales beyond toy problems.
- → Co-trained model + harness design signal production-grade autonomous engineering workflows.