Highlights

Meta has released Muse Code (beta), a terminal‑based coding agent driven by the Muse Spark 1.2 coding model and available through the Meta Model API and a curl install script. The agent coordinates persistent subagents, logs every event to a local replay‑exact event store, and can resume after a crash without losing progress. On Meta’s benchmark charts Muse Spark 1.2 trails Anthropic’s Opus 5 on shown coding tests but generally outperforms OpenAI’s Codex and Google’s Antigravity. The standout feature is the restart‑safe runtime: long jobs can resume exactly where they left off, which matters more than raw speed for extended engineering workflows.

Sentiment Analysis

• Overall sentiment is mixed to slightly positive. The release is notable for practical engineering features rather than outright benchmark dominance. The progress bar below reflects a cautiously positive tone, highlighting useful innovations but acknowledging stronger competitors in some tests.
  
  
  

  60%

  
  
  
  

Article Text

Meta has introduced Muse Code (beta), a terminal‑based coding agent built on the Muse Spark 1.2 model and distributed via the Meta Model API and a simple curl installation. The product is aimed at software engineering across large repositories: planning changes, writing code, validating outcomes, and coordinating multiple persistent subagents to tackle complex problems with less human intervention. Meta positions this release as an incremental but meaningful step forward while indicating that larger models are on the way.

A key operational detail of Muse Code is its event logging strategy. The agent records every model call, tool invocation, approval, and edit to a local event log that functions as a single source of truth. Because of this design, the runtime is "replay‑exact" and restart‑safe: if the process crashes, it can pick up precisely where it left off. For long‑running tasks that may span hours or days, this resilience can be more important than raw inference speed, and it is a distinguishing capability compared with many competing agents.

Muse Code ships with built‑in skills and commands that support iterative and approval‑gated workflows. For example, a "/plan" command converts a task into a plan requiring approvals; "/grill" stress‑tests a plan until it proves robust; and "/goal" drives toward completion of the objective. Meta reports that Muse Spark 1.2 was co‑trained with Muse Code so the underlying LLM and the agent’s orchestration logic operate in synergy.

On benchmarks, Muse Spark 1.2 tells a nuanced story. It is a coding‑focused update over the previous generation, with Meta saying it increased compute for code tasks and broadened the training environments to improve code generation, debugging, and end‑to‑end workflows. In Meta’s reported results, Muse Spark 1.2 with Muse Code scored below Anthropic’s Opus 5 on Terminal‑Bench 2.1 (82.9% vs. 86.7%) but ahead of several other systems such as GPT‑5.6 Terra on Codex and Grok Build in the same charts. Agent‑oriented benchmarks like DeepSWE 1.1 showed smaller gaps: Muse was closer to competitors though still behind Opus 5 in the reported numbers.

Performance over longer sequences and many tool calls is another important dimension. Meta’s speedup charts indicate Opus 5 achieves the largest gains over baseline across many tool calls, while Muse Spark 1.2 occupies a middle position. Meta emphasizes that Muse Code’s behavior improves as tool calls accumulate — the characteristic you want from a long‑horizon coder that refines solutions over time.

Some of the most compelling demos are long‑horizon and multimodal. Meta demonstrated Muse Code iteratively optimizing GPU kernels across more than 1,000 tool calls over extended runs on Nvidia Hopper hardware, showing the agent could improve outputs over prolonged work. There are also multimodal examples: a user supplied a fly‑through video of a home (an mp4) and Muse Code interpreted the video to generate a visually rich website with booking functionality. Such demos reflect Meta’s multimodal ambitions across the Muse family.

Despite these strengths, Meta enters a crowded market. Other companies have shipped agentic coding tools and parallel agent orchestration; some established alternatives already offer comparable or broader capabilities. Meta’s differentiator is not outright benchmark supremacy but the crash‑safe runtime and subagent architecture. That design could be attractive for teams that run long, autonomous workloads, though it also raises the usual concerns about powerful agents operating with extended autonomy and repeated tool access.

In short, Muse Code (beta) is a pragmatic release that emphasizes robustness for long‑running developer workflows and multimodal capabilities. It does not unseat the current leaders on every benchmark, but it introduces features — especially the replay‑exact event log and persistent subagents — that may change how teams approach extended, agented engineering tasks. The agent is available for testing now via a straightforward install command provided by Meta.

The practical takeaway: Muse Code’s restart‑safe runtime and coordinated subagents make it particularly suited to long‑horizon, multimodal engineering jobs where continuity and reproducibility are more valuable than single‑task speed.

Key Insights Table