Meta Platforms has introduced Muse Code, a terminal-based coding agent powered by its recently released AI model, Muse Spark 1.2.
The tool is aimed at assisting software developers with writing and debugging code.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
Muse Code has been designed for developers working on large, complex software projects.
The agent can not only generate code and verify its accuracy, but also manages several persistent background sub-agents that operate throughout each session.
Meta says this capability is intended to reduce redundant information gathering and speed up the handling of difficult tasks by allowing these background agents to independently progress and determine when to report back to the main agent.
The company stated Muse Code’s design incorporates a local event log, recording every model call, tool use, approval, and edit.
This system is intended to allow the tool to recover precisely from failures, enabling it to tackle tasks that may require prolonged processing times without losing progress in case of interruptions.
Among its functions, Muse Code offers default commands such as /plan to turn assignments into approval-gated plans, /grill to stress-test those plans, and /goal to work towards set objectives.
The tool supports interpreting a variety of user inputs, including video files, and can generate outputs such as tailored marketing and booking pages for holiday properties.
Muse Spark 1.2, which underpins the new coding agent, is described by Meta as an update to Muse Spark 1.1, with improvements in code generation, debugging, codebase comprehension, and support for end-to-end developer workflows.
The new version involved expanded training on large coding tasks and a range of programming environments.
Meta also reported that the model was co-trained with Muse Code to optimise performance when used together.
In evaluations, Meta tested Muse Spark 1.2 on tasks such as the optimisation of GPU kernels, benchmarking the agent’s ability to iteratively improve performance using its coding environment.
According to the company, Muse Spark 1.2 was able to develop optimised solutions for specific Nvidia GPU workloads with its built-in planning and code generation features.
Muse Spark 1.2 is now available for use both within Muse Code and via the Meta Model API. This latest release arrives around a month after Meta introduced Muse Spark 1.1 for developer testing.
The earlier version was used to generate coding challenges that informed improvements in Muse Spark 1.2, particularly in following complex instructions.
Meta first launched the Muse Spark series in April as its entry into a staged model development approach. The architecture now drives the latest features in the Meta AI assistant and related platforms.
