Meta's AI research division has entered the agentic coding race with Muse Code, a new terminal-based coding agent, and Muse Spark 1.2, the upgraded model that powers it. The release, announced August 5, 2026, marks Meta's most direct attempt yet to challenge dedicated coding tools from OpenAI, Anthropic, and a wave of startups. For developers following the latest AI tools and agentic platforms, Muse Code offers a distinctly research-driven take on how an AI pair programmer should behave.
Muse Code is designed to take on complex software engineering tasks across large codebases — planning changes, writing code, and validating the results. Meta says the agent represents its "next step toward the frontier," with larger and more capable models on the way.
A Multi-Agent Design Built for Long Tasks
The most notable architectural choice in Muse Code is its use of persistent background agents. Rather than spawning a fresh subagent for each subtask, Muse Code keeps a set of asynchronous background agents alive throughout an entire session. These agents gather information and carry out next steps on their own, reporting back to the main agent only when relevant.
Meta argues this design reduces latency and the need for constant human steering on difficult, multi-step problems. "Their persistence reduces latency and the need for steering," the company wrote, because the agents avoid redundantly re-gathering context they already hold.
Muse Code is also built to survive crashes. Every model call, tool execution, approval, and edit is appended to a local event log that serves as a single source of truth. That makes the runtime "replay-exact and restart-safe," according to Meta — after a failure, the agent can resume precisely where it stopped, enabling it to tackle long-running tasks without being derailed.
The agent ships with several built-in skills: a `/plan` command that turns a task into an approval-gated plan, a `/grill` command that stress-tests that plan until it holds up, and a `/goal` command that works toward a specified objective.
Muse Spark 1.2: A Coding-First Upgrade
Muse Spark 1.2 is a coding-focused update to the previous Muse Spark 1.1 model. Meta said it improved the model's code generation, complex debugging, codebase understanding, and end-to-end developer workflows by significantly scaling up training compute on coding tasks and expanding the diversity of its training environments, while maintaining its strengths in general agent tasks.
Crucially, Meta co-trained Muse Spark 1.2 alongside Muse Code itself. The training incorporated rejection-sampled harness trajectories and recipe optimizations for goals, context compaction, and subagents, along with integration of the Muse Code toolset to maximize compatibility.
The company also employed a self-improvement loop: it used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates, then had the model grade candidate solutions on how well they met the requirements. That process produced a scalable training dataset that helped Muse Spark 1.2 follow complex instructions more precisely than its predecessor.
Benchmarking on GPU Kernel Optimization
To demonstrate Muse Spark 1.2's capabilities, Meta tested the model's ability to iteratively optimize GPU kernels over more than 1,000 tool calls, with sessions lasting up to 24 hours. Using Muse Code's agentic environment, the model wrote, compiled, profiled, and progressively improved kernel performance against a provided baseline.
The benchmarks targeted KDA and MLA kernels running on NVIDIA Hopper GPUs. Meta reported that the agent achieved "substantial improvements" over the baseline implementations. For the MLA workload, the model designed a two-kernel Triton pipeline combining kernel fusion and tiling with optimizations specific to the latent-attention architecture. Models were barred from simply importing third-party kernel libraries and had to implement the algorithms from scratch in Triton.
Availability and What's Next
Muse Spark 1.2 is available today through Muse Code and via the Meta Model API, with what Meta describes as "expanded global access." The company was emphatic that this is only an interim step, teasing "new harness features and more powerful models" on the horizon.
The launch places Meta squarely in a crowded and competitive field. Coding agents have become one of the hottest categories in AI, with tools promising to autonomously handle everything from bug fixes to entire feature implementation. Meta's differentiator is its deep research pedigree and its emphasis on persistent, crash-resilient agents that can run unattended for hours.
Whether developers adopt Muse Code over entrenched alternatives remains to be seen, but the signal from Meta is clear: the company intends to compete seriously in agentic software engineering, not just consumer chatbots. With larger models promised soon, the coding-agent wars are set to intensify.
Stay Ahead of AI
New AI tools and coding agents are shipping every week. To track the latest AI developments — from model releases to developer platforms — keep AI Buzz Wire in your rotation.
Read more AI news →
