Meta Launches Muse Code, Testing Long-Running Coding Work With 1,000+ Tool Calls

Muse Code can be tested against a practical hypothesis over the next three to six months: a capable repository-scale agent should shorten pull-request cycle time while reducing human rework and keeping the total inference cost below the labor it saves. Meta has so far published its own evaluations and demonstrations. If enterprise trials do not improve those three measures, the launch will not yet show that an agent can reliably take responsibility for complete engineering tasks.

Meta Superintelligence Labs released the Muse Code beta on August 5, 2026. It is a terminal-based coding agent powered by Muse Spark 1.2 and currently supports macOS and Linux. Meta says the agent can inspect large repositories, plan changes, write code, validate results, and coordinate persistent background agents during a session. 9to5Mac independently confirmed the supported platforms, beta status, and installation route, while Reuters reported the product and model release the same day.

An event log is designed to make interrupted work recoverable

Muse Code writes model calls, tool use, approvals, and file edits to a local event log. That record allows a task to resume after a crash instead of restarting every expensive step. Built-in commands can produce a plan, stress-test that plan, or continue working toward a stated goal. The mechanism addresses two familiar constraints in long-running agent work: context compaction can discard earlier decisions, and a failed process can otherwise erase operational state. The log improves traceability, but Meta’s announcement does not quantify recovery success across repository sizes or set limits for token use, latency, and cost when several background agents run together.

Meta says it co-trained Muse Spark 1.2 with Muse Code using agent trajectories and optimizations for goals, context compaction, and subagents. The company also used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates, after which the model graded candidate solutions. This self-improvement loop can expand training data, but the generator, grader, and evaluated system remain within Meta’s process. External teams still need to repeat the comparisons with their own repositories, test suites, and review rules.

The 24-hour case study focuses on GPU kernel optimization

In Meta’s published case study, the system performed 1,000+ tool calls over as long as 24 hours while repeatedly writing, compiling, profiling, and improving GPU kernels. The work covered KDA and MLA kernels on NVIDIA Hopper GPUs. For KDA, models could not directly import third-party kernel libraries; for MLA, Meta compared against a PyTorch reference implementation at batch size 1, 64 heads, sequence length 8192, and latent dimension 512. Those disclosed conditions make the experiment more inspectable. They still describe two controlled optimization tasks and cannot establish performance for ordinary enterprise debugging, access control, or deployment across services.

Muse Spark 1.2 is available through Muse Code and the Meta Model API, for which Meta says global access has expanded. The release post does not provide a general-availability date, per-task cost, enterprise administration details, or third-party benchmark results for Muse Code. The most informative next evidence will be total time to close the same issue, number of human interventions, test pass rate, and the resulting bill in external trials. A long-running agent becomes a procurement-ready engineering tool only if those measures improve together.

Sources:

• Meta AI Research: Introducing Muse Code and Muse Spark 1.2
• Reuters: Meta launches new AI coding tool powered by Muse Spark 1.2
• 9to5Mac: Meta launches Muse Code AI coding agent for macOS and Linux