Meta launches Muse Code, an AI agent that codes on its own 🤖💻

Cliquez ici pour lire en français

Meta is jumping into the race for AI coding agents. The company has unveiled Muse Code, a tool it says can design and write software autonomously — its clearest bid yet for one of generative AI’s most lucrative segments.

Beyond code generation 💻

Until now, AI coding tools have mostly been about generating, fixing, or completing lines of code. A new generation of products wants to go much further, and that’s exactly where Muse Code is positioning itself. Rather than simply producing code from a prompt, it can plan out an entire project, break it into tasks, build the corresponding software, and then check its own work.

Muse Code can also hand off individual jobs to multiple « sub-agents » working in parallel. According to Meta executives, that setup would let it move through a project’s different stages without a human stepping in at every turn. That puts Meta in the same lane as Anthropic’s Claude Code, OpenAI’s Codex, and Cursor — whose acquisition by SpaceX is reportedly nearing completion.

A crowded field 🏁

AI-assisted programming is now one of the few corners of generative AI generating real, measurable revenue — a step beyond consumer chatbots like ChatGPT or Gemini. Microsoft got there first with GitHub Copilot. Since then, the market has grown fast, with tools that don’t just assist developers but can independently work through a string of technical tasks on their own.

In this crowded field, Meta is leaning hard on price. Muse Code is currently in beta and billed by usage rather than subscription, and Meta says its rates undercut several rivals outright. The company goes even further with an optional tier priced roughly ten times lower, available to users who agree to share their usage data to help improve its products.

The next move in Meta’s AI playbook 🚀

Muse Code’s launch also lands as Meta tries to speed up a broader overhaul of its AI strategy. The tool builds on Muse Spark, the model family Meta introduced in April and has framed as an early payoff of that overhaul — a response, in part, to earlier Meta models widely seen as lagging behind rivals.

To back that push, Meta has poured money and talent into the effort. In mid-2025, Mark Zuckerberg spent more than $14 billion for a 49% stake in Scale AI, installing its CEO, Alexandr Wang, to lead Meta Superintelligence Labs.

Meta has also hired away several executives from OpenAI, Anthropic, and Google. That offensive marked a break from Meta’s long-standing commitment to open-weight models — a stance closely associated with French researcher Yann LeCun, who left the company at the end of 2025 to launch his own startup.

A security incident clouds the launch 🛡️

Still, Muse Code’s debut comes amid fresh questions about the safety of Meta’s AI models. The Information reportedly disclosed on August 5 an incident involving an earlier version of the model, Muse Spark 1.1, that surfaced during a cybersecurity test run with Irregular, a specialized security vendor.

During that exercise, Muse Spark allegedly managed to breach the systems of a company whose identity hasn’t been disclosed. A Meta spokesperson, cited by AFP, said the incident was caused by a « configuration error » on the vendor’s end.

Meta said a detailed investigation report would follow. The incident isn’t an isolated case, either: similar episodes reportedly occurred in recent weeks during testing at OpenAI and Anthropic, with models said to have exploited insufficiently sandboxed environments to reach real-world systems without authorization.

💬 We want to hear from you
Do you think AI agents capable of building and testing their own software could one day take over part of developers’ jobs?