Meta launches a new AI coding agent designed for repository-scale development, debugging, and long-running engineering workflows
Meta has expanded its artificial intelligence portfolio with the launch of Muse Code (Beta), a new terminal-based AI coding agent powered by its latest Muse Spark 1.2 model. Designed specifically for software engineers, the new platform aims to simplify complex development tasks by helping developers plan, write, debug, validate, and manage code across large-scale repositories while supporting extended engineering workflows.
Currently available in beta for macOS and Linux, Muse Code represents Meta’s latest push into AI-assisted software engineering, focusing on improving productivity for developers working on complex projects that require sustained reasoning over extended periods.
A Coding Agent Built for Modern Software Engineering
Unlike conventional AI coding assistants that primarily generate code snippets, Muse Code is built to serve as an intelligent software engineering partner capable of handling complete development workflows.
The AI agent can create structured coding plans, implement code changes, verify outputs, and coordinate multiple background AI agents that continuously work across extensive codebases. This persistent approach enables developers to delegate repetitive engineering tasks while maintaining visibility and control over every step of the development process.
According to Meta, the release marks another milestone in its broader roadmap for building increasingly capable AI systems for programming.
“We’re excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model. This marks our next step towards the frontier, with larger and much more capable models on the way,” the company said in its announcement.
Persistent AI Agents Replace Traditional Session-Based Assistants
One of Muse Code’s defining features is its persistent agent architecture.
Instead of launching a new AI assistant for every coding request, Muse Code maintains long-running background agents throughout an entire development session. These agents continuously gather context, execute assigned tasks, and coordinate with one another, reducing the amount of manual intervention required from developers.
The platform also maintains a local event log, recording every AI model interaction, tool execution, user approval, and code modification. This design enables interrupted sessions to resume seamlessly without losing progress, making the platform particularly useful for long-running software engineering projects.
Built-In Commands Simplify Development Workflows
Muse Code introduces several integrated commands aimed at making software development more structured and predictable.
The /plan command converts user requests into an execution strategy that requires approval before implementation. Developers can then use /grill to evaluate and stress-test the proposed plan before any code changes are made.
Another feature, /goal, allows developers to assign broader engineering objectives, enabling the AI system to progressively work toward completing larger tasks instead of responding only to individual prompts.
These workflow-oriented tools are intended to move AI coding beyond simple code generation and toward comprehensive project execution.
Muse Spark 1.2 Brings Major Coding Improvements
At the core of Muse Code is Muse Spark 1.2, Meta’s latest coding-focused language model and an upgrade over Muse Spark 1.1.
The company says the new model significantly improves several critical software engineering capabilities, including:
- Code generation
- Debugging
- Repository-level code understanding
- End-to-end software development
- Long-duration engineering workflows
These improvements stem from increased training compute dedicated specifically to coding tasks, along with exposure to a wider variety of software development environments.
Muse Spark 1.2 is available both through Muse Code and via the Meta Model API, allowing developers and enterprises to integrate the model into their own applications and workflows.
Co-Trained With Muse Code for Better Collaboration
Meta explained that Muse Spark 1.2 and Muse Code were developed together to improve compatibility between the AI model and the coding agent.
The company incorporated optimisation techniques such as rejection-sampled harness trajectories, enhanced planning strategies, context compaction, and improved coordination between multiple AI subagents. Muse Code’s own toolset was also integrated directly into the training process, enabling tighter interaction between the model and the development environment.
This co-training approach allows the coding assistant to better understand complex workflows while maintaining context over lengthy development sessions.
Designed for Long-Horizon Software Projects
A major focus of Muse Spark 1.2 is handling long-horizon coding tasks; projects that extend far beyond isolated programming requests.
Meta says the model was trained on large-scale software engineering scenarios involving complete repositories, extensive software projects, and automated research workflows.
To sustain performance across these lengthy tasks, the model employs advanced planning techniques, goal conditioning, and context compaction, enabling it to retain relevant information and continue making progress even during extended coding sessions.
AI Helping Train the Next Generation of AI
Meta also revealed that Muse Spark 1.1 played a direct role in building its successor.
The earlier model was used to generate coding environments, instruction-following templates, and candidate solutions that were subsequently evaluated and incorporated into the training process for Muse Spark 1.2. The company described this iterative methodology as a self-improvement loop, allowing one generation of AI to help create more capable future models.
Tested on Complex GPU Optimisation Tasks
To evaluate the model’s capabilities, Meta subjected Muse Spark 1.2 to demanding GPU kernel optimisation challenges.
According to the company, the AI completed more than 1,000 tool calls across development sessions lasting up to 24 hours. During these tests, the model independently wrote, compiled, profiled, and refined NVIDIA Hopper GPU kernels, ultimately delivering measurable performance improvements over baseline implementations without relying on third-party kernel libraries.
The results demonstrate the model’s ability to sustain complex reasoning and iterative optimisation across extended engineering workflows.
Strengthening Meta’s AI Developer Ecosystem
With the introduction of Muse Code and Muse Spark 1.2, Meta is strengthening its position in the rapidly evolving AI coding landscape. Rather than focusing solely on code generation, the company is targeting full software engineering workflows by combining persistent AI agents, repository-scale reasoning, and long-duration task execution.
As AI continues to reshape software development, Meta’s latest offerings signal a shift toward intelligent engineering assistants capable of collaborating with developers throughout the entire lifecycle of large and complex software projects.
Read more:Â 10-Year-Old Japanese Student Jo Nagai Discovers Butterflies Remember Life as Caterpillars
