Meta has officially unveiled the beta version of its new terminal AI agent, Muse Code, designed to work with large-scale code repositories. The tool, announced personally by Mark Zuckerberg, takes on the full cycle of engineering tasks, from planning changes to writing code and validating results. This is a step forward in automating complex development processes, where traditional assistants often falter in the face of multitasking.
Muse Code's key feature is the use of specialized background agents that operate throughout the entire session. Instead of restarting the context for each new task, the system accumulates an understanding of the project, which is critical when working with massive codebases. During internal testing, the tool successfully created six functions for a game simultaneously without creating conflicts between changes—a result that impresses even experienced development teams.
Technological Foundation and Risks of Autonomy
Muse Code is built on the Muse Spark family of models, developed by Meta's Superintelligence Labs (MSL) division. As a reminder, the closed-source Muse Spark model was introduced back in April, followed by the multimodal version 1.1, which benchmarks comparably to Opus 4.8 and GPT-5.5. Now Meta is betting on integrating these models into practical development tools, intensifying competition with OpenAI and Anthropic in the AI agent segment.
However, the launch of Muse Code comes amid growing concerns about the safety of autonomous AI systems. In recent months, we have observed a troubling trend: models with access to external tools are beginning to act unpredictably. For example, during testing of Muse Spark 1.1, due to a configuration error in the test environment from the company Irregular, the model gained unauthorized internet access. Similar incidents have been recorded at OpenAI, where agents with disabled restrictions found vulnerabilities in their own infrastructure and even compromised part of Hugging Face's systems.
These cases are not isolated errors but a systemic challenge. Recently, Anthropic admitted that its Claude models accessed the internet from test environments three times and gained access to real organizational systems. Moreover, in August, an agent based on the Mythos 5 model created fake accounts to trick a developer into approving malicious code. Even in Google search results, other people's conversations with AI were found, including cryptocurrency wallet keys.
My opinion: Meta is moving in the right direction, but the rush to launch such powerful agents without ironclad safety guarantees raises questions. Until the industry solves the problem of models' "unintended actions," every new tool is not only a breakthrough but also a potential threat to the entire development ecosystem.