Meta releases Muse Code for persistent AI software engineering

Meta has launched Muse Code, a beta AI coding agent specifically built to navigate and manage complex software assignments within expansive code repositories. This tool utilizes the new Muse Spark 1.2 model to provide persistent assistance for developers working on macOS and Linux platforms during difficult development cycles.

Persistent agent architecture and reliability

The core innovation of this platform lies in its use of specialized background agents. Unlike traditional AI tools that create a new instance for every individual prompt, these agents stay active during the entire work session. They operate asynchronously and choose the most appropriate moments to transmit their findings to the primary interface. This approach minimizes the need for developers to provide constant direction or repeat context for multi-step projects.

Reliability is a central focus for this release. Meta integrated a local event log that records every specific tool run, model call, and code edit. Because every action is appended to this log, the runtime remains restart-safe. If the system experiences a crash, the agent can resume its work exactly where it left off without losing progress. This makes the tool suitable for long-running tasks that might take hours to complete.

The Muse Spark 1.2 model is the engine behind these capabilities. It is available through the Muse Code interface and the broader Meta Model API. Meta has also moved to expand global access to this API to reach a wider audience of developers. By keeping agents persistent, the system avoids the information gathering delays that often plague non-persistent AI coding assistants.

This architecture supports more fluid workflows. Developers can assign a high-level task and allow the background agents to handle the granular details. The primary agent acts as a coordinator, while the background workers perform the heavy lifting. This structure is intended to mirror how human engineering teams operate on large-scale software projects.

Model training and performance benchmarks

Meta took a unique approach by co-training the Muse Spark 1.2 model alongside the Muse Code agent. This process ensured that the model was optimized specifically for the workflows and tools provided in the software environment. The training data included full-repository generation and end-to-end software projects to prepare the model for the scale of modern enterprise applications.

Despite these efforts, some industry analysts believe co-training is becoming a standard practice. Rivals like OpenAI and Anthropic are already treating the engineering of the agent harness as a core part of the training cycle. The true test for Meta will be whether this integration leads to fewer human interventions on real-world projects. If the model can plan and handle context better than its peers, it will gain a significant foothold.

Performance data shows a competitive landscape. On the Terminal-Bench 2.1 evaluation, the model achieved a pass rate of 82.9 percent. This placed it slightly ahead of the GPT-5.6 Terra model but behind Claude Opus 5. However, on the DeepSWE 1.1 benchmark, the model scored 59.3 percent, which trailed behind both of its primary competitors.

Meta pointed out that these evaluations used specific agents for each model. This means that a model might perform better or worse depending on the prompts and tools designed for it. Analysts suggest that the most important benchmark for any technology officer will be how the tool performs against their own internal company pipelines. Real-world utility in a private environment often differs from standardized scores.

Hurdles for enterprise implementation

While the technology is impressive, companies face significant obstacles before they can fully adopt persistent AI agents. Security and governance remain at the top of the list for most IT departments. Many organizations are hesitant to connect AI agents to their continuous integration and continuous delivery environments. These systems are sensitive, and a mistake by an autonomous agent could cause significant downtime.

Identity management is another concern. Enterprises need to ensure that AI agents follow the same access protocols as human employees. This requires deep integration with existing security systems to prevent unauthorized access to sensitive parts of a codebase. Without these controls, the risk of data leaks or unauthorized code modifications is too high for many large corporations.

Cost forecasting presents its own set of challenges. It is difficult for managers to predict how many tokens a persistent agent will consume during a complex task. Meta offers two different pricing tiers to address different needs. The lower-priced Contributor tier costs $0.10 per million input tokens but allows Meta to use the data to improve its products. The standard tier is more expensive at $1.25 per million input tokens but offers better data privacy.

Finally, there is the issue of vendor lock-in. If a company builds its entire development workflow around a specific AI agent, it becomes harder to switch to a different provider later. This reliance can limit long-term flexibility. Most experts believe that initial adoption will focus on low-risk tasks before agents are trusted with critical production code. Only after proven success will companies allow these tools to operate with higher levels of autonomy.