# Meta joins the AI coding race with Muse Code and Muse Spark 1.2 Meta's new coding assistant stands out with background agents and parallel work capabilities. - **Source:** [VentureBeat](https://venturebeat.com) - **Published:** 2026-08-05 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/b2ce4e29-823e-4d45-b7ef-847dcc268124 ## Meta's new coding assistant Meta has launched Muse Code, a terminal-based AI coding assistant in beta. This tool can handle full software engineering tasks across large code repositories: planning changes, writing code, and validating results. Muse Code can be installed on macOS or Linux with a single curl command and is offered as a completely proprietary product. Meta CEO Mark Zuckerberg hinted on X about a possible open-source version in the future. ## Entering the competition Muse Code represents Meta's most serious entry into the AI-assisted coding space. The company now directly competes with rivals like Anthropic's Claude Code and OpenAI's Codex. While Meta previously focused on open-weight Llama models, Muse Code marks a strategic shift toward proprietary products. This move could spark significant competition in the AI coding tools market. ## Background agents technology The standout feature of Muse Code is what Meta calls 'async background agents.' Unlike competing tools that spawn new agents for each task, Muse Code maintains specialized background agents throughout the entire session. This approach eliminates redundant information gathering and offers developers lower latency. ## Parallel work capability Muse Code uses parallel sub-agents for large-scale tasks. Each sub-agent works in its own isolated git worktree, ensuring the developer's working copy remains untouched. Zuckerberg shared on X that they successfully built six game features simultaneously without collisions using this feature. This capability could significantly boost productivity for teams working with large codebases. ## Auditability feature Another key feature is Muse Code's ability to log all model calls, tool runs, approvals, and edits to a local event log. This makes the runtime exactly replayable and restart-safe. If Muse Code crashes during a long-running task, it can resume precisely where it left off without lost work or re-prompting - a major advantage for enterprise users. ## Bundled skills Muse Code comes with integrated skills familiar to users of competing tools. The /plan command converts tasks into approval-gated plans, /grill stress-tests these plans until they're robust, and /goal directs the agent toward completing specified objectives. These built-in skills give developers more controlled management of the AI assistant. ## Muse Spark 1.2 features Underlying Muse Code is Muse Spark 1.2, a coding-focused AI model update. Meta enhanced it with significantly increased training compute for coding tasks and broader training environment diversity. The update improves code generation, complex debugging, and codebase understanding while maintaining general agent capabilities, specifically targeting the Muse family's previous weaknesses. ## Model and harness co-training A notable aspect of Muse Spark 1.2 is its co-training with the Muse Code harness. Meta used rejection-sampled harness trajectories and optimizations for goals, context compaction, and sub-agents to tune the model specifically for this tool. This approach reflects an industry-wide shift away from treating models and harnesses as separate products. ## Self-improvement loop Meta employed a self-improvement loop to train Muse Spark 1.2. Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions against these requirements to create scalable training data for its successor. Meta credits this loop with measurably improving 1.2's ability to follow complex instructions. ## Benchmark results Meta published benchmark charts comparing Muse Spark 1.2 against other coding models on Terminal-Bench 2.1, DeepSWE 1.1, and an internal Meta coding benchmark. Notably, the announcement didn't highlight any rankings - unusual reticence in a field where competitors trumpet leaderboard victories. Meta directed readers to a separate methodology report for details. ## Market impact Muse Code's launch could cause significant ripples in the AI-assisted coding tools market. Meta's entry offers enterprise customers, especially those working with large codebases, a new option. Competitors are expected to respond to innovative features like auditability and background agents. This development signals AI's growing role in software development processes. ## Future outlook Meta's launch of Muse Code and Muse Spark 1.2 signals the company's long-term strategy in AI-assisted coding. Zuckerberg's hints about possible open-sourcing may be testing community reaction. In coming months, adoption rates and real-world performance will determine Meta's position in this space. As AI assistants become increasingly critical for developers, intensified competition seems inevitable. ## Sources covering this story - https://venturebeat.com/orchestration/meta-enters-the-ai-coding-wars-with-muse-spark-1-2-and-muse-code-with-persistent-async-background-agents — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._