Meta launches Muse Code beta, crash-resilient coding agent

Meta released Muse Code (beta), a terminal coding agent powered by the Muse Spark 1.2 model and made it available through the Meta Model API and a curl install script. The agent records a replay-exact local event log so it can restart precisely after crashes. Meta reported immediate availability for testing.

The runtime records every model call, tool run, approval and edit to a single local event log. According to Meta, that log acts as a single source of truth and allows the agent to resume from the point of interruption without re-running prior steps. The runtime design is presented as a feature for long-running engineering tasks.

Muse Code is aimed at software engineering across large repositories. In its announcement Meta described the agent as able to “plan changes, write code, and validate the results.” The company says Muse Code can coordinate persistent background subagents to work on tasks with less human intervention.

The agent includes built-in commands and skills. The “/plan” command turns a requested task into an approval-gated plan, “/grill” stress-tests that plan until it holds up, and “/goal” drives toward completion of the objective. Meta said it co-trained Muse Spark 1.2 with Muse Code so the model and runtime are aligned.

Meta published benchmark comparisons showing mixed results. On Terminal-Bench 2.1, Muse Spark 1.2 with Muse Code scored 82.9 percent, behind Anthropic’s Opus 5 at 86.7 percent and ahead of OpenAI’s Codex at 81.8 percent and Grok Build at 81.6 percent. On DeepSWE 1.1, focused on agentic coding capabilities, Muse scored 59.3 percent versus Opus 5 at 65.0 percent and Codex at 64.8 percent. Meta’s internal coding benchmark recorded Muse at 70.6 percent compared with Opus 5 at 79.4 percent.

Meta also published speedup charts over more than 1,000 tool calls. Those charts show Opus 5 with the largest gains versus baseline, about 74–75 percent, while Muse Spark 1.2 registered improvements in the roughly 61–69 percent range depending on the run. Meta presented these charts as measures of how agent performance changes as tool calls accumulate.

Meta demonstrated long-horizon and multimodal use cases. In stress tests Muse Code iteratively optimized GPU kernels over more than 1,000 tool calls in runs lasting up to 24 hours on Nvidia Hopper GPUs. In a separate multimodal demonstration a user supplied a fly-through MP4 of a house and the agent produced a website with booking capabilities after interpreting the video.

The agent can be installed with the command provided by Meta: curl -fsSL https://dev.meta.ai/install.sh | bash. Meta positioned Muse Code’s crash-resilient runtime and persistent subagent design as distinguishing features and noted that larger, more capable models are planned for future releases.