Muse Code: Meta's crash-resilient bet in the coding agent race

Meta spent a decade building social apps and a reputation in AI research. That reputation came from spending billions to give the research away. Now it is shipping a terminal coding agent. Muse Code, in beta for macOS and Linux, installs with a single command, curl -fsSL https://dev.meta.ai/install.sh | bash, and pairs with a new model, Muse Spark 1.2, that Meta co-trained with the agent's own tool harness.

The agent takes on engineering work across large repositories: planning changes, writing code, validating results. For each task it can coordinate multiple persistent subagents, and those subagents stay alive for the whole session rather than being spawned per job. Meta says that cuts redundant information gathering and reduces how much steering a developer has to do on difficult, multi-step tasks.

A runtime built to survive its own failures

The event log is the design detail that matters. Muse Code appends every model call, tool run, approval, and edit to a local log that becomes the single source of truth for the session. The runtime is replay-exact and restart-safe: if the agent crashes, it resumes precisely where it stopped instead of starting over. That is what lets it take on long-running tasks without being derailed by failures.

Meta ships three default commands with the agent. /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal works toward completing the stated objective. Structuring work as an explicit procedure is the same idea behind research on treating SOPs as code, which found that compiling a procedure into executable form alone lifts strong agents by 16 points. The company's demo shows another side of the tool: it accepts an mp4 fly-through of a home as input and produces a vacation-home marketing and booking page from it.

Muse Spark 1.2, co-trained with its own harness

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. Meta scaled up training compute on coding tasks and expanded training environment diversity, while keeping the model's strength in other areas like general agents.

The unusual part is how it was trained. Meta co-trained Muse Spark 1.2 with Muse Code itself, using rejection-sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, plus integration of the Muse Code toolset. The training also covered long-horizon coding: whole-repository generation, large end-to-end projects, and auto-research, with planning, goal conditioning, and context compaction to sustain progress.

Then there is the self-improvement loop. Muse Spark 1.1 generated challenging coding environments and instruction-following templates; the model then graded candidate solutions against those requirements, producing a scalable training dataset for the 1.2 release. That is how the newer model learned to follow complex instructions more precisely than its predecessor.

The 24-hour kernel test

Meta's case study pushes the long-horizon claim hardest. The agent iteratively optimized GPU kernels over more than 1,000 tool calls, running up to 24 hours. It writes, compiles, profiles, and progressively improves kernel performance relative to a provided baseline, benchmarked on KDA and MLA kernels for NVIDIA Hopper GPUs.

The model was barred from importing third-party kernel libraries such as FLA directly; it had to apply its own optimization knowledge to implement the algorithms in Triton rather than wrapping existing code. Meta reports that the agent achieved substantial improvements over the baseline in both workloads, without publishing exact figures in the announcement.

Where Meta fits in the agent race

Muse Code lands as AI coding tools decouple from the physical keyboard. Cursor's cloud agents let developers dispatch tasks asynchronously and review results from anywhere, as Cursor's SQLite rebuild showed. Apple's Xcode 27 brings agents from Anthropic, Google, and OpenAI directly into the IDE. On the model side, Anthropic's Claude Sonnet 5 is built for agentic workflows and enterprise use. The field is converging on one question: which tool can be trusted to keep working without a human at the keys.

Meta's answer is infrastructure. A replay-exact event log, persistent background agents, and a model trained alongside its own harness all point at the same goal: letting an agent run for hours without supervision and pick itself up after a failure. That is a different bet than competing on a single benchmark number, at a moment when the old benchmarks are saturated and contaminated.

Availability and what's next

Muse Spark 1.2 is available today in Muse Code and through the Meta Model API with expanded global access. Meta frames the release as "our next step toward the frontier" and says larger, much more capable models are on the way.

The demo video is a good party trick, and the kernel case study is worth a read. The durable claim is operational: an agent that survives crashes and needs less steering is the one developers will trust with overnight jobs, the trust problem Circle's agent-stack research flags as the hard half. Meta's bet is that this race is won on reliability, and that it can get there by shipping to a terminal near you.