[DigitalToday reporter Jinju Hong] Meta has unveiled a beta version of Muse Code, a terminal-based artificial intelligence (AI) coding agent, stepping into the race to automate AI software development. It appears to be trying to differentiate itself from OpenAI, Anthropic and others by emphasizing long autonomous execution and crash recovery rather than a battle for top performance.

On Aug. 5 (local time), blockchain media outlet Decrypt reported that Meta launched Muse Code in beta, an AI coding agent based on its latest coding model, Muse Spark 1.2.

Muse Code can be used immediately through Meta's model API and an installation script. It is designed to handle the full software engineering process, including making change plans in large code repositories, writing code and verifying results. Meta said it was "happy to release the Muse Code beta" and that it was also preparing "bigger and much more powerful models."

Meta put the most emphasis on its execution structure. Muse Code records model calls, tool executions, user approvals and code edit histories in a local event log. It is designed to manage task status based on the log so that, even if a system failure occurs, it can resume work precisely from where it stopped. Meta explained that "even after a crash, the agent can restart exactly from the point where it stopped."

It also provides features for long development tasks. It supports basic commands such as "/plan" to create a work plan, "/grill" to repeatedly review the plan and "/goal" to continue working until the objective is completed. Meta said it co-trained Muse Spark 1.2 so that the large language model (LLM) and the agent can collaborate organically.

On benchmark performance, it did not reach the top tier. According to data Meta released, Muse Spark 1.2 scored lower than Anthropic's Opus 5 on all major coding benchmarks. It outperformed OpenAI's Codex and Google's Antigravity in most categories.

On Terminal Bench 2.1, which evaluates terminal-based coding capability, Muse Code scored 82.9 percent. In the same category, Claude Code, based on Opus 5, scored 86.7 percent, while Codex scored 81.8 percent and Grok Build scored 81.6 percent.

On DeepSWE 1.1, which measures an AI agent's software development capability, Muse scored 59.3 percent, Opus 5 scored 65.0 percent and Codex scored 64.8 percent. On Meta's internal coding benchmark, Muse scored 70.6 percent, below Opus 5 at 79.4 percent.

Meta highlighted stability in long-running work rather than simple performance. The company said that in long-run environments where tool calls exceed 1,000, Muse Spark 1.2's performance improved to about 61 to 69 percent depending on execution conditions. In a live demonstration, it repeatedly optimized GPU kernels by making more than 1,000 tool calls over 24 hours on Nvidia Hopper GPUs.

In another demo, the agent analyzed drone footage a user shot inside a home and automatically created a website that included a booking function. Meta explained that it "interpreted the video and generated a visually rich booking website."

By the timing of market entry alone, Meta is a latecomer. OpenAI already offers Codex, a parallel cloud-based coding agent, and DeepSeek has also introduced a competing product to Claude Code. Open-source AI coding tools such as Hermes and OpenClaw are also spreading quickly. Meta is presenting crash recovery, long autonomous execution and the operation of subordinate agents as differentiators rather than competing for top performance.

The challenge remains that AI agents that operate autonomously for long periods can raise productivity while reducing predictability.

Meta sees this structure as the direction needed for future development environments and has first released a beta version. Developers can test Muse Code directly using only the installation command.