Meta has introduced Muse Code, a terminal-based coding agent powered by Muse Spark 1.2, designed to plan, write, and validate code across large repositories and sustain multi-step engineering work with reduced user intervention
Meta has released Muse Code, a terminal-based software agent intended to automate extended and complex software engineering workflows. Built on the Muse Spark 1.2 model, Muse Code is designed to operate across large codebases, planning and executing multi-step changes, validating results, and coordinating persistent background agents to manage follow-up tasks. Unlike conventional code-generation tools that produce isolated snippets, Muse Code aims to sustain direction over long sequences of engineering decisions, reducing the need for repeated user input and minimizing latency in multi-stage development processes.
In a series of internal tests, Muse Code completed more than 1,000 tool calls during sessions lasting up to 24 hours. One demanding evaluation required the system to optimize GPU kernels for Nvidia Hopper hardware, using KDA and MLA kernels as test cases. Rather than importing third-party libraries, Muse Code-via Muse Spark 1.2-iteratively wrote, compiled, profiled, and refined kernel code to outperform a provided baseline. The model applied algorithm-specific optimizations, including chunk-parallel preparation, sequential inter-chunk scanning, fusion, and tiling. These tests were designed to assess whether the agent could maintain coherent progress across extended engineering tasks, rather than simply generating a correct answer in a single pass.
Muse Code's architecture includes persistent background agents that remain active throughout a session, carrying out delegated work and determining when to report results to the main agent. This approach is intended to reduce redundant information gathering and lower the frequency of user intervention. The system also maintains a local event log that records model calls, tool executions, approvals, and edits, enabling what Meta describes as a replay-exact and restart-safe runtime. If a session is interrupted, Muse Code can resume from the last recorded state, which is particularly relevant for workflows involving repeated compilation, testing, and optimization where late-stage failures could otherwise result in significant lost work.
To support extended workflows, Muse Code provides built-in tools such as /plan for approval-gated planning, /grill for stress-testing plans, and /goal for working toward specified objectives. The Muse Spark 1.2 model was co-trained with Muse Code, incorporating agent trajectories and tool-environment interactions into its training data. Meta reports that Muse Spark 1.2 is also being used for broader software engineering tasks, including whole-repository generation, complex debugging, and long-running projects. The model is accessible through Muse Code and Meta's Model API, with plans for expanded features and larger models in future releases.
In another demonstration, Muse Code was shown generating a vacation-home marketing and booking page from an MP4 fly-through video, illustrating its ability to process non-traditional inputs beyond source code. This reflects a broader trend in AI agent development, where systems are increasingly expected to handle multimodal data and coordinate complex tool use. Comparable advances have been reported elsewhere in the field, such as Google DeepMind's recent expansion of Gemini Robotics for full-body humanoid control, highlighting the growing ambition to automate multi-step technical work across domains.
During the reported tests, Muse Code executed over 1,000 tool calls in sessions lasting up to 24 hours, with persistent agents managing background tasks and a local event log supporting crash recovery. The system was evaluated on GPU kernel optimization tasks using Nvidia Hopper GPUs, with performance measured against a supplied baseline rather than third-party libraries. These figures are developer-reported, and independent verification of sustained reliability and generalization to diverse engineering environments remains outstanding.
Understanding agent-based coding systems like Muse Code requires attention to the distinction between automated code generation and sustained, multi-step engineering workflows. While conventional code-generation models typically produce isolated outputs in response to prompts, agentic systems are designed to plan, execute, and validate sequences of actions, often coordinating multiple tools and maintaining state across extended sessions. This introduces new challenges in reliability, error recovery, and human oversight, as well as opportunities for reducing repetitive manual work in large-scale software projects. The effectiveness of such agents depends not only on model quality but also on robust session management, tool integration, and transparent logging to support auditability and safe recovery from failure.
