Meta debuts Muse Code agent

Meta on Wednesday introduced Muse Code, its first AI coding agent, opening a new front in its competition with Anthropic and OpenAI. The tool is in beta and is powered by Muse Spark 1.2, Meta’s latest coding-focused model, according to the company and CNBC. (research.meta.ai) Meta said Muse Code is designed to handle full software engineering tasks across large code repositories, including planning changes, writing code and validating results. CNBC reported the launch as Meta’s latest push under AI chief Alexandr Wang, who leads Meta Superintelligence Labs. (cnbc.com) ### What exactly did Meta release on August 5? Meta said Muse Code is a terminal-based coding agent that developers can install on macOS or Linux with a single command. (research.meta.ai) The company described it as a beta product built for “complex software engineering tasks across large repositories,” rather than a chat interface limited to code snippets. CNBC reported that Muse Code is Meta’s first coding agent and that the company is positioning it against Anthropic and OpenAI. Alexandr Wang told CNBC the product can be used to “take on complete software engineering tasks across a wide variety of use cases, planning changes, writing code, validating the results.” (research.meta.ai) ### How is Muse Code supposed to work inside a developer workflow? Meta said Muse Code uses a main agent plus persistent background agents that stay active through a session. Those subagents can handle next steps and feed information back to the main agent, which Meta said reduces latency and cuts down on repeated information gathering during multi-step tasks. (research.meta.ai) Mark Zuckerberg said in a social media post, as reported by TechCrunch, that the system can split larger jobs into separate subagents working in parallel in isolated worktrees. He said the setup leaves a developer’s working copy untouched, and he said Meta had tested it by building six game features simultaneously “with no collisions.” (cnbc.com) ### What is new about Muse Spark 1.2? Meta said Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1 with improvements in code generation, debugging, codebase understanding and end-to-end developer workflows. The company said it scaled up training compute on coding tasks and expanded the diversity of training environments. (research.meta.ai) The company also said Muse Spark 1.2 was co-trained with Muse Code so the model would perform better when paired with the coding agent. CNBC reported Wang said the joint development of the model and the tool improved coding performance. (techcrunch.com) ### How is Meta pricing the product? CNBC reported developers can access Muse Code through a pay-as-you-go option with pricing similar to the Muse Spark 1.1 release: $1.25 per million input tokens and $4.25 per million output tokens. Wang also told CNBC there is a lower-cost “contributor tier” that is more than 10 times cheaper than the pay-as-you-go tier if developers opt in to help improve the model. (research.meta.ai) TechCrunch reported Wang separately described the product as competitive on cost for many workflows. That pricing pitch places Muse Code in direct comparison with Anthropic’s Claude Code and OpenAI’s Codex-style coding tools, which CNBC and other outlets cited as the main rivals. (research.meta.ai) ### Why does this matter inside Meta’s broader AI push? CNBC reported Muse Code is the latest major release overseen by Wang at Meta Superintelligence Labs. The outlet said Wang joined last year as part of Chief Executive Mark Zuckerberg’s effort to rebuild Meta’s AI strategy and that the company is looking for more ways to generate revenue from AI while continuing to spend heavily on data centers and computing infrastructure. (cnbc.com) Meta’s own announcement pointed to larger and “more capable” models still to come. For now, the next concrete step is the beta rollout of Muse Code on macOS and Linux, with developers able to access it through Meta’s install flow and pay-as-you-go tiers tied to Muse Spark 1.2. (cnbc.com) (research.meta.ai)