What is Meta's new Muse Code AI agent, what capabilities does it offer for large code base tasks, how does it use parallel sub agents and Muse Spark, and how does it position Meta competitively against coding agents from OpenAI and Anthropic?

Meta launched Muse Code on August 5, 2026 — a terminal-based AI coding agent in beta, powered by the new Muse Spark 1.2 model, that directly competes with Anthropic's Claude Code and OpenAI's Codex . It is designed to handle complete software engineering tasks across large repositories, from planning changes to writing code and validating results .

Large codebase capabilities

• End-to-end engineering tasks: Plans changes, writes code, runs tests, and validates results across large repos in a single session .
• 1M-token context window — enough to hold large codebases in context simultaneously .
• Isolated git worktrees for each subagent, enabling safe parallel experimentation without conflicts .
• Restart-safe execution — sessions survive crashes, preserving task history and audit logs .
• Full audit logging: every action by every subagent is recorded for transparency .
• Bundled design skills (UX, architecture patterns) rather than requiring external prompts .
• Installs with a single terminal command on macOS and Linux .

Parallel sub-agents and Muse Spark

• Automatic subagent fan-out: When Muse Code receives a task, it spawns multiple persistent sub-agents that work in parallel on different parts of the problem .
• Persistent, async background agents — unlike tools that create agents per step, Muse Code's subagents remain active throughout the session and coordinate continuously .
• Muse Spark 1.2 was co-trained alongside the Muse Code agent harness itself, meaning the model was optimized for the multi-agent, long-horizon workflow from the ground up .

Competitive positioning

Meta is competing primarily on price, crash resilience, and architecture, not raw benchmark superiority:

• Aggressive pricing: Muse Code is pitched as one of the most affordable coding agents on the market . It offers a standard pay-as-you-go tier (~$1.25/M input tokens) and a cheaper "contributor" tier where Meta can use prompts for training .
• Terminal-Bench 2.1 score of 82.9% — placing it between Claude Code and Codex in Meta's reported evaluation .
• Key differentiators: persistent background agents (vs. per-step agents in competitors), isolated worktrees for safe parallel work, crash-safe sessions, and a model co-trained with the agent rather than adapted afterward .
• Meta is not claiming outright benchmark leadership — instead, it is positioning Muse Code as a lower-cost, more resilient alternative for large, long-running engineering tasks .