Mark Zuckerberg Unveils Muse Code AI to Automate Software Engineering Tasks
Meta CEO Mark Zuckerberg has introduced Muse Code, a new artificial intelligence coding agent designed to handle complete software engineering tasks. Unlike traditional AI assistants that primarily help developers write individual lines of code, Muse Code is built to manage entire software projects with minimal human intervention. The tool has been launched in beta and is powered by Muse Spark 1.2, Meta's latest coding-focused AI model.
The launch comes amid an accelerating race among technology companies to develop smarter coding assistants capable of automating software development. Muse Code is positioned to compete directly with existing products in this space, including Anthropic's Claude Code and OpenAI's Codex.
Multi-Agent Architecture and Persistent Memory
Announcing the beta launch on the social media platform X, Zuckerberg described Muse Code as a terminal coding agent capable of executing comprehensive software engineering tasks across large repositories. According to the Meta CEO, the tool can plan changes, write code, and validate the resulting output.
A key feature of Muse Code is its persistent memory structure. While many AI assistants start each new task without any memory of previous interactions, Muse Code maintains specialised AI agents that run continuously throughout a developer's session. Zuckerberg explained that these background agents gradually develop a deeper understanding of the project as they work, enabling them to manage larger and more complex codebases over time.
For larger assignments, Muse Code can distribute the workload across multiple AI sub-agents. Each sub-agent operates independently within its own isolated workspace, ensuring that a developer's existing code remains untouched while multiple tasks are executed simultaneously. To demonstrate this capability, Meta tested the system by directing it to build six game features at the same time, completing the tasks without any collisions between the active AI agents.
Reliability Features and Performance Benchmarks
Meta is positioning Muse Code as a reliable tool for long-running development projects. The system records every model request, tool call, and code edit in a local event log before execution. This logging mechanism ensures that if the system crashes midway through a task, it can resume from the exact point of interruption rather than requiring the developer to restart the process.
Zuckerberg also shared performance results from an internal benchmark where Meta applied Muse Spark 1.2 to a kernel optimisation task running on NVIDIA Hopper hardware. During a 24-hour period, the model made more than 1,000 tool calls and continued to identify performance improvements long after its initial exploration phase.
Currently, Muse Code can be installed with a single command and is accessible through Meta's contributor tier. Zuckerberg indicated that the release of Muse Spark 1.2 is an initial step, noting that larger and more capable AI models are already in development at the company.
The Shifting Landscape of Software Engineering
The introduction of Muse Code coincides with bold predictions from industry leaders regarding the future of software development. Earlier this year, Boris Cherny, the creator of Anthropic's Claude Code, suggested that the role of the software engineer will undergo rapid changes as AI agents become capable of managing nearly every phase of the development process.
Speaking during a talk at Anthropic, Cherny stated that there is a strong possibility developers will stop using traditional integrated development environments (IDEs), such as Microsoft's VS Code and Apple's Xcode, by the end of the year. He previously suggested that all software engineering roles will begin to become obsolete starting this year.
According to Cherny, Anthropic chose to build Claude Code as a terminal-based tool because AI models are advancing so rapidly that developing traditional graphical interfaces may no longer be practical. Cherny claimed that Claude Code can already write complete features, fix bugs, and navigate large codebases with minimal human guidance, adding that he has not manually changed a line of code since November 2025, relying instead on AI agents to perform the work.
