**Meta Enters the Coding Agent Race with Muse Code (Beta)**
Meta has officially entered the competitive arena of AI-powered coding agents with the release of **Muse Code (beta)**. Powered by its latest **Muse Spark 1.2** model, the new terminal-based agent represents the company’s latest push to challenge established leaders like OpenAI and Anthropic in the developer tool space.
At its core, Muse Code is designed to function as a persistent, autonomous partner for software engineering. Unlike simpler chatbots, it operates as a “terminal coding agent,” capable of managing complex, long-running tasks across large code repositories. Its key differentiator is a sophisticated runtime environment that logs every single action—be it a model call, a tool execution, or a user approval—into a single, replay-exact event log. This architecture makes the system “restart-safe”; if a process crashes, the agent can theoretically resume execution from the exact moment of failure, a significant advantage for reliability.
The agent is equipped with default “skills” activated via simple commands. The `/plan` command helps structure complex tasks, `/grill` stress-tests the proposed solutions, and `/goal` works iteratively toward completing the objective. Meta has co-trained the underlying Muse Spark 1.2 model with the agent itself, aiming for deep synergy between the core LLM and the agentic layer.
### Performance and Benchmarks
Performance data from Meta’s own charts reveals a competitive landscape. While Muse Spark 1.2 does not top every benchmark, it holds its ground against formidable competitors. On the Terminal-Bench 2.1 evaluation, it achieved a score of **82.9%**, trailing behind Claude Code (Opus 5) at **86.7%** but outperforming OpenAI’s GPT-5.6 Terra and xAI’s Grok Build.
DeepSWE 1.1, a benchmark focused on agentic coding capabilities, told a similar story, with Muse scoring **59.3%** compared to Claude’s **65.0%** and Codex’s **64.8%**. Meta’s internal benchmarks showed Muse hitting **70.6%** against Opus 5’s **79.4%**.
However, Meta emphasizes that Muse Code’s true strength lies in its long-horizon execution. In stress tests, the agent demonstrated the ability to iterate over **1,000+ tool calls for up to 24 hours**, continuously improving its output. Furthermore, Muse Code showcases multimodal capabilities; in a recent demo, the agent successfully interpreted a fly-through video of a house and generated a fully functional, visually rich website with booking features.
### The Competitive Landscape
Despite the ambitious features, Meta arrives in a crowded market. OpenAI’s Codex already powers parallel cloud agents, and DeepSeek has launched its own rival to Claude Code. Agentic tools like Hermes and OpenClaw are also mature alternatives. Meta’s primary edge, therefore, is not raw benchmark performance but the unique combination of a crash-safe runtime and a multi-subagent design that promotes persistence and accuracy over time.
**FAQ**
**What is Muse Code, and how does it work?**
Muse Code is a terminal-based coding agent powered by Meta’s Muse Spark 1.2 model. It is designed to handle complex software engineering tasks by planning changes, writing code, and validating results. It operates by coordinating persistent background subagents and maintaining a detailed event log, allowing it to resume exactly where it left off after a crash.
**What makes Muse Code different from other coding agents?**
Its standout feature is its “replay-exact” runtime. Every action is logged in a single source of truth, making the agent restart-safe. It also comes with built-in skills like `/plan` for structuring work and `/grill` for stress-testing ideas.
**How does it perform compared to competitors?**
While it trails behind leaders like Claude Opus 5 on most benchmarks, it performs competitively on tasks like DeepSWE and demonstrates a significant advantage in long-horizon, iterative problem-solving over extended sessions.
**How can I try Muse Code?**
Muse Code is available for testing. Users can install it by running a specific curl command provided by Meta.
**What is the target use case for Muse Code?**
It is built for developers who need an autonomous partner for large-scale software engineering, particularly for projects that require sustained effort across multiple files and complex dependencies.
### Conclusion
Meta’s Muse Code (beta) marks a significant milestone in the democratization of agentic coding. By prioritizing reliability and persistent execution over sheer benchmark scores, Meta has positioned Muse Code as a robust tool for developers who value stability in long-running tasks. While the battle for AI coding supremacy is far from won, Muse Code’s unique replay-safe architecture ensures that it is a contender worth watching in the rapidly evolving landscape of AI-assisted software development.
