Meta Enters Coding Agents With Muse Code and Muse Spark 1.2 — Async Subagents, Replay-Safe Runtime
Meta has entered the terminal coding agent market with Muse Code, a beta release powered by Muse Spark 1.2. It is Meta’s direct play for the developer tooling segment now dominated by Claude Code and OpenAI Codex.
The Agent Architecture
Muse Code’s distinguishing design choice is persistent async background agents. Where most coding agents spawn subagents for individual tasks and tear them down on completion, Muse Code keeps specialized background agents alive for the full session. Each remains active, continues working on next steps, and communicates back to the main agent only when relevant. The result: less repeated context gathering, lower latency on multi-step tasks.
The runtime uses a local event log. Every model call, tool run, approval, and edit is appended in order. This makes the runtime replay-exact and restart-safe — if the agent crashes mid-task, it can resume from exactly where it stopped. That capability makes 24-hour unattended runs practical.
Three built-in skills ship with the agent: /plan converts a task into an approval-gated execution plan, /grill stress-tests that plan until it holds, and /goal drives the agent toward a defined objective.
Muse Spark 1.2
The underlying model was co-trained with the Muse Code harness itself — rejection-sampled harness trajectories, recipe optimizations for goals, context compaction, and subagent coordination were all included in training. This tight coupling between model and scaffold mirrors the approach Anthropic has used for Claude Code.
Training was scaled up on long-horizon coding tasks: whole-repository generation, large end-to-end projects, and auto-research workflows. The model uses goal conditioning to maintain direction over extended runs and context compaction to preserve working knowledge without exceeding context limits.
A self-improvement loop was used: Muse Spark 1.1 generated challenging coding environments and instruction-following templates. The new model then graded candidate solutions, creating a scalable training dataset for its successor.
Benchmark Position
Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index, placing it at #13 globally across 185 evaluated models. Pricing is $1.25 per million input tokens and $4.25 per million output tokens — cache hit price drops to $0.15 per million. Context window is 1 million tokens.
At 54 on the AA Index, Muse Spark 1.2 sits in the same tier as Kimi K2.6 and Xiaomi MiMo-V2.5-Pro. It is meaningfully behind Claude Opus 5 (max, 65+), Fable 5 (64.9), and GPT-5.6 Sol (max), but competitive with Grok 4.5 (62) and roughly level with GLM-5.2.
Kernel Optimization Case Study
Meta published one internal benchmark: GPU kernel optimization. Muse Spark 1.2, operating via Muse Code with 1,000+ tool calls over up to 24 hours, iteratively optimized KDA and MLA kernels for NVIDIA Hopper GPUs. The model paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, applying KDA-specific optimizations — re-centering the gated cumulative decay at the chunk midpoint — rather than wrapping existing libraries. Meta reports “substantial improvements over the provided baseline implementation.”
The benchmark is vendor-reported and specific to one use case. External SWE-bench data for Muse Spark 1.2 is not yet available.
Availability
Muse Code is available in beta at dev.meta.ai. Muse Spark 1.2 is available through the Meta Model API with expanded global access. Meta describes this as “a next step toward the frontier, with larger and more capable models on the way.”
The coding agent market now has five credible terminal-native entries: Claude Code (Anthropic), OpenAI Codex, Cursor Composer, xAI Grok Build, and Muse Code. Each is co-developed with its underlying model — the era of drop-in model swaps for coding agents appears to be ending.