The landscape of software development tools has shifted steadily toward systems that can manage extended sequences of work rather than isolated snippets.
Developers routinely face repositories that span thousands of files, interdependent modules, and requirements that unfold over hours or days of iteration. Even for the most tech-savvy and the battle-hardened, this can be a pain, since it can take much of their time.
Major AI labs have solutions for this.
They have created specialized agents capable of sustained planning, execution, and verification without constant human redirection.
Now, it's Meta's turn.
The company known for its massive portfolio of messaging apps and social media platforms, have released 'Muse Code.'
Muse Code is essentially a beta terminal-based agent.
It runs a basic agent loop supported by a group of asynchronous background agents that stay active for the entire session instead of being created for single tasks. These agents handle subsequent steps, decide when to report back to the main agent, cut down on repeated information collection, lower latency, and reduce the amount of human steering needed on complex multi-step work.
The system records every model call, tool execution, approval, and edit in a local append-only event log.
That log acts as the sole source of truth, making the runtime fully replayable and able to restart exactly where it left off after a crash, which supports long-running tasks without interruption from failures.
Default skills included with Muse Code are /plan, which converts a task into an approval-gated plan; /grill, which repeatedly stress-tests the plan until it is robust; and /goal, which drives work toward successful completion of a stated objective.
These abilities are powerted by an updated model called 'Muse Spark 1.2,' which allows the agent to handle complex, multi-file changes across large codebases by first outlining a plan, then implementing the necessary edits, and finally validating the outcomes.
It relies on a set of persistent asynchronous background agents that remain active for the duration of a session rather than being created and discarded for each subtask.
These background processes carry out intermediate steps, limit repeated information gathering, and lower the amount of steering required on multi-step problems.
The underlying runtime records every model invocation, tool execution, approval, and edit in an append-only local event log.
That log serves as a single source of truth, allowing the system to replay sequences exactly and to resume after interruptions or crashes without loss of state.
Default skills shipped with the agent include commands for converting a task into an approval-gated plan, stress-testing that plan, and driving progress toward a stated objective until it is met.
Muse Spark 1.2 itself is a coding-oriented revision of the earlier Muse Spark 1.1 checkpoint.
"In Muse Spark 1.2, we significantly scaled up training compute on coding tasks while expanding training environment diversity. The model also maintains its strength in other key areas like general agents," said Meta in a blog post.
Training compute directed at coding tasks was increased substantially, and the diversity of training environments was broadened.
The resulting gains appear in code generation, complex debugging, codebase comprehension, and complete developer workflows, while general agentic capabilities are described as remaining intact.
The model was co-trained alongside the Muse Code harness, incorporating trajectories sampled from the agent, optimizations for goal conditioning and context compaction, and direct integration of the agent’s tool set.
Additional emphasis was placed on long-horizon activities such as whole-repository generation, large end-to-end projects, and automated research sequences that rely on sequential planning and selective retention of context.
One concrete evaluation involved iterative optimization of GPU kernels.
Over more than a thousand tool calls spanning periods of up to 24 hours on Nvidia's Hopper hardware, the agent produced Triton implementations for KDA and MLA kernels that achieved competitive performance improvements relative to supplied baseline versions.
Separate demonstrations illustrate multimodal input handling: a short fly-through video of a residential interior supplied as an MP4 file was interpreted by the agent, which then generated a full marketing and booking website reflecting the visual details of the space.
Both the agent and the model are available immediately.
Muse Code is available through a one-line installation script for macOS and Linux, whereas Muse Spark 1.2 can be accessed through the agent itself as well as the Meta Model API, which has been extended with broader geographic access.
The release reflects a broader shift in AI-assisted software development.
Rather than serving as autocomplete for code, systems like Muse Code are being designed to function as persistent collaborators that can plan, execute, verify, and recover from interruptions over extended periods.
As repositories continue to grow in size and complexity, that ability may become just as important as raw coding performance.
With Muse Code, Meta joins the growing group of companies building long-running software engineering agents, including OpenAI's Codex, Anthropic's Claude Code, and Google's Antigravity CLI.
Now, with Meta in the race, the competition to build AI that can independently tackle large-scale engineering tasks is accelerating.
