Meta has launched Muse Code, a new AI coding agent aimed at enhancing software engineering tasks. The tool features advanced logging and subagent capabilities.
Summary
- Meta has launched Muse Code (beta), a terminal-based coding tool utilizing the upgraded Muse Spark 1.2 model, accessible via the Meta Model API.
- This innovative coding agent features persistent background subagents and a detailed event log, enabling it to resume operations precisely from the point of interruption.
- Meta's evaluations indicate that Muse Spark 1.2 falls short of Anthropic's Opus 5 in coding benchmarks but surpasses OpenAI's Codex and Google's Antigravity in various aspects.
Meta has entered the competitive arena of AI coding agents with the introduction of Muse Code (beta), positioning itself against established players like Anthropic and OpenAI.
In an official announcement, Meta expressed excitement about the launch, stating, "We are thrilled to unveil Muse Code (beta), powered by our latest Muse Spark 1.2 model," highlighting its significance in their future development plans.
Muse Code is tailored for software engineering tasks within expansive repositories, facilitating complex operations such as planning, coding, and outcome validation. Meta claims it can manage multiple persistent subagents for each task, improving efficiency and accuracy while minimizing the need for human oversight.
A key feature of Muse Code is its runtime capability, which meticulously logs every model call, tool execution, and modification, ensuring that the agent can recover seamlessly after a crash. This feature is particularly valuable for lengthy tasks, offering a reliability that competitors have yet to emphasize.
Additionally, Muse Code incorporates several default commands, including "/plan" for task planning, "/grill" for rigorous testing, and "/goal" for guiding task completion, designed to work in harmony with its core LLM.
Performance Metrics
Muse Spark 1.2 represents a significant upgrade from its predecessor, Muse Spark 1.1. Meta has reported enhancements in training compute and diversity in the training environment, leading to improvements in code generation and debugging capabilities.
In recent benchmarks, Muse Spark 1.2 achieved a score of 82.9% in Terminal-Bench 2.1, trailing behind Opus 5's 86.7% but outperforming Codex and Grok Build. In agentic coding capabilities measured by DeepSWE 1.1, Muse scored 59.3%, compared to Opus 5's 65.0% and Codex's 64.8%.
Meta's internal coding benchmark showed Muse achieving 70.6%, while Opus 5 scored 79.4%. Notably, Muse demonstrated adaptability over extended tasks, showcasing improvements in performance as tool calls increased, which is beneficial for long-duration coding tasks.
In multimodal demonstrations, Muse Code illustrated its ability to convert a fly-through video of a house into a visually engaging website, aligning with Meta's vision for the Muse series.
Here’s an example of Muse Code’s multimodal visual coding capabilities. In this demo, the user inputs a fly-through video of a home into the terminal as an mp4 file. Muse Code interprets the video and produces a visually rich website with booking capabilities. pic.twitter.com/3CAfIMYmAB
— AI at Meta (@AIatMeta) August 5, 2026
Competitive Landscape
Despite its innovative features, Meta's entry into the coding agent market comes after competitors like OpenAI and DeepSeek, which have already established their own coding agents with substantial capabilities. Muse Code's unique strengths lie in its crash-safe operation and subagent architecture rather than benchmark performance.
The unpredictable nature of agentic coding poses inherent risks; however, Meta is optimistic that developers will appreciate the autonomy offered by Muse Code. The tool is currently available for testing via a simple installation command.
