Product signal
Meta launches Muse Code coding agent
Meta has launched its beta terminal coding agent, Muse Code, alongside Muse Spark 1.2, entering the coding-agent product race with OpenAI and Anthropic.
Meta has launched Muse Code, its first beta terminal coding agent, alongside Muse Spark 1.2. The key change is not simply another model release: Meta is packaging model capabilities into a developer product that can execute complex programming tasks and delegate work to sub-agents, entering a workflow layer where OpenAI and Anthropic already compete.
From model previews to an agent product
Meta describes Muse Code as a beta terminal coding agent released with Muse Spark 1.2. Several media outlets frame it as Meta’s first coding-agent offering. That differs from Muse Spark 1.1, which Meta released through a public preview of the Meta Model API in July. The developer touchpoint is moving from calling a model to assigning work in a terminal. Users will therefore judge not only generated code, but whether the agent can finish bounded engineering tasks.
The mechanism is task orchestration
The Information reports that Muse Code can delegate complex work to sub-agents, allowing developers to split programming tasks among multiple AI agents. Meta had previously said Muse Spark 1.1 supported tool use, computer use, complex codebases, planning, and context compression. Muse Code turns those capabilities into a terminal workflow. The contest is consequently about harnesses, permissions, context management, task decomposition, and feedback loops—not just single-turn code quality.
The new capability still needs independent testing
Meta’s developer page lists an 80.0 Terminal-Bench 2.1 score for Muse Spark 1.1, versus 82.7 for Opus 4.8 and 83.4 for GPT-5.5; those are Meta-presented results. Antigma measured 76.2%±1.23 for Muse Spark 1.1 and 78.9%±1.31 for Claude Code with Opus 4.8, but did not test Muse Code or Muse Spark 1.2. Axios had also reported that Meta’s agents lagged leading products in task breadth and autonomous runtime.
What to watch next
Observable next evidence includes Meta’s disclosure of Muse Code API scope, pricing, regional availability, and enterprise controls; reproducible third-party tests of Muse Code and Muse Spark 1.2 on Terminal-Bench, SWE-bench, or real repositories; and signs that developers integrate it into CI, IDE, or agent-orchestration workflows. Persistent gaps in independent completion rates or long-running reliability would weaken the competitive claim.
Sources
- CNBC — Meta debuts first AI coding agent to take on Anthropic and OpenAI
- Bloomberg Technology — Meta Debuts AI Coding Agent in Race With OpenAI and Anthropic
- The Information — Meta Debuts New AI Coding Tools
- The Wall Street Journal — Meta Releases Coding Agent to Compete with OpenAI and Anthropic
- Meta — Introducing Muse Code and Muse Spark 1.2
- Meta Superintelligence Labs — Introducing Muse Spark 1.1
- Meta — Products & Solutions for AI Developers
- Axios — Meta's Muse Spark agents connects to Google Calendar and Gmail
- Simon Willison — Introducing Muse Code and Muse Spark 1.2
- The Decoder AI — The company that made open weights mainstream now competes on discounts
- Artificial Analysis — Muse Spark Intelligence, Performance & Price Analysis
- Meta AI Research — Muse Spark 1.2 & Muse Code Evaluation Methodology