Meta is introducing Muse Code, a terminal coding agent that runs on the new Muse Spark 1.2 model. The agent plans code changes, writes code, and validates the results within large repositories. Meta promises that larger and more powerful models are on the way.
The tool tackles complex engineering tasks in large codebases. According to Meta, the agent coordinates multiple persistent subagents per task, allowing difficult problems to be solved faster and with less human intervention.
A new feature is that Muse Code works with asynchronous background agents. These remain active throughout a session, rather than being launched per task. This eliminates the need to repeatedly gather the same information. The agent uses a local event log that records every model call, tool run, and operation. If the system crashes, the agent resumes exactly where it left off.
Muse Spark 1.2 as the engine
Under the hood, Muse Spark 1.2 runs, an update to version 1.1 specifically trained for coding. Meta significantly scaled up the training compute resources for coding tasks and expanded the diversity of training environments. The model was co-trained with Muse Code to ensure optimal alignment between the model and the agent.
Muse Spark 1.2 focuses explicitly on long-horizon tasks, such as generating complete repositories and large end-to-end projects. Meta also used a self-improvement loop: its predecessor, Muse Spark 1.1, generated challenging coding environments and then evaluated the solutions.
Muse Spark 1.2 is available worldwide starting today in Muse Code and via the Meta Model API. Meta reports that new harness features and more powerful models are in the pipeline.
Tip: Agents from OpenAI and Anthropic use social engineering to carry out hacks