Meta has launched Muse Code in beta, giving developers a coding agent designed to take on larger software engineering jobs rather than simply suggesting the next few lines of code.
Powered by the new Muse Spark 1.2 model, Muse Code can plan changes across large repositories, write code and validate its own results. It can also coordinate several persistent subagents at the same time, allowing parts of a more complicated development task to be handled in parallel.
That puts Meta more directly into a rapidly expanding market for autonomous coding tools, where Anthropic’s Claude Code and OpenAI’s Codex are already competing for developers. The direction is becoming fairly clear: coding assistants are being asked to do increasingly less assisting and considerably more of the actual work.
AWS has been pursuing much the same idea with its Kiro autonomous agent, which it positions as a virtual developer capable of working autonomously for extended periods. Muse Code approaches the problem slightly differently, but both reflect the growing expectation that an AI coding system should be able to take responsibility for a task rather than requiring a developer to continually move it forward.
Muse Code keeps its agents running
A large part of Meta’s approach involves background agents that remain active throughout a Muse Code session.
Rather than creating a new subagent every time additional work is required, Muse Code keeps specialised agents running in the background. They can gather information, carry out subsequent steps and decide when information needs to be sent back to the main agent.
The idea is to avoid repeatedly rebuilding context during complicated jobs, while reducing the amount of steering required from the developer.
That becomes increasingly important as coding agents are expected to work across entire repositories and complete tasks that might involve dozens or hundreds of individual actions. Maintaining useful context over that kind of workload is considerably harder than generating a function or fixing an isolated error.
Meta has also built Muse Code around a local event log that records every model call, tool invocation, approval and edit. If the agent crashes, it can use that history to resume from where it stopped rather than reconstructing the task from scratch.
For software engineering work that could continue for hours, that reliability is less glamorous than another benchmark number, but probably more consequential.
Meta is building workflows around the agent
Muse Code ships with several built-in skills intended to structure how it tackles a job.
The /plan command creates an approval-gated plan before work begins, while /grill is designed to stress-test that plan before execution. A /goal skill gives the system an objective and allows it to keep working towards successful completion.
These features give some indication of where Meta thinks coding agents need to improve. Generating code is only part of the problem once an AI system is given responsibility for a larger project. It also has to decide what to do next, retain the objective across a long series of actions and recognise when its initial approach needs to change.
Muse Code is built around Muse Spark, the AI model family Meta introduced in April through Meta Superintelligence Labs. The original model already supported tool use and multi-agent orchestration, while Muse Spark 1.1 expanded its coding and agentic capabilities when it arrived in July.
Muse Spark 1.2 pushes further into software development, with Meta saying it has increased the amount of training compute dedicated to coding while expanding the range of training environments used by the model.
Muse Spark 1.2 and Muse Code were trained together
Instead of building the coding interface around an independently developed model, Meta says Muse Spark 1.2 and Muse Code were co-trained.
The training included trajectories generated through the coding environment itself, along with optimisation around goals, context compaction and the use of subagents. Meta says this helps the model work more effectively with the tools and workflow available inside Muse Code.
Long-running tasks were another focus. Muse Spark 1.2 was trained on jobs including whole-repository generation, end-to-end software projects and automated research, where maintaining useful context over a long sequence of actions becomes critical.
Meta also used Muse Spark 1.1 to help create more difficult coding environments and instruction-following tasks for its successor. The earlier model graded candidate solutions against those requirements, producing additional training data for Muse Spark 1.2.
It is a slightly recursive development process: Meta is using one generation of its AI system to create some of the conditions used to train the next.
Meta tested it for more than 1,000 tool calls
One of Meta’s examples involved asking the system to optimise GPU kernels over more than 1,000 tool calls, with individual runs lasting for as long as 24 hours.
During the process, Muse Code repeatedly wrote, compiled and profiled kernels before trying to improve their performance against a supplied baseline. Meta tested the system on KDA and MLA kernels running on Nvidia Hopper GPUs and says it continued finding performance improvements as the task progressed.
A company-run demonstration obviously doesn’t tell us how reliably Muse Code will perform across an unfamiliar production codebase, particularly when a task is poorly defined or the consequences of an incorrect change are significant. It does, however, show what Meta is trying to optimise for.
The useful measure for this generation of coding agents is becoming less about whether they can produce convincing code from a prompt and more about whether they can keep making sensible decisions after the 10th, 100th or 1,000th action.
Muse Code is available in beta on a pay-as-you-go basis, while Muse Spark 1.2 is also being made available through the Meta Model API. Meta’s standard Muse Code tier is priced at $1.25 per million input tokens and $4.25 per million output tokens.
The coding-agent market is moving quickly enough that simply having one is unlikely to differentiate Meta for long. What will matter is whether Muse Code can maintain the context, judgement and reliability required when developers stop using these systems as clever autocomplete and start giving them meaningful portions of a software project to complete.