Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by the New Muse Spark 1.2 Model Meta AI 发布 Muse Code（测试版）：由全新 Muse Spark 1.2 模型驱动的终端编码智能体

TL;DR

• Meta AI released Muse Code (beta), a terminal coding agent for macOS/Linux powered by the new Muse Spark 1.2 model, targeting complex software engineering across large repositories
• Muse Spark 1.2 was co-trained with the Muse Code harness itself, using rejection-sampled trajectories, long-horizon repository-scale training, and self-improvement loops from Muse Spark 1.1-generated environments
• The agent architecture features persistent async background agents that remain alive throughout sessions (avoiding per-task spawning overhead) and an append-only local event log that is replay-exact and restart-safe after crashes
• A notable case study demonstrated iterative GPU kernel optimization over 1,000+ tool calls spanning up to 24 hours on NVIDIA Hopper GPUs, producing optimized KDA and MLA Triton kernels
• Evaluation on Terminal-Bench 2.1, DeepSWE v1.1, and Meta Internal Coding Bench (440 tasks from real PRs) positions Muse Spark 1.2 as a frontier coding model with expanded global API access

Why It Matters

Meta's simultaneous release of a coding-focused model and its dedicated agent harness represents a significant step toward production-grade autonomous software engineering, demonstrating that co-training models with their operational tools yields measurable performance gains. The persistent agent architecture and crash-resilient event logging address two of the most persistent challenges in long-running AI coding workflows, making this directly relevant to teams building or adopting agentic development pipelines.

Technical Details

• Architecture: Muse Code runs a main agent loop plus persistent async background agents that stay alive across sessions, eliminating redundant information gathering. Three bundled skills ship by default: /plan (approval-gated task planning), /grill (stress-testing plans), and /goal (objective-driven execution)
• Training methodology: Muse Spark 1.2 was co-trained with the Muse Code harness using rejection-sampled trajectories, recipe optimizations for goals/compaction/subagents, and long-horizon training on whole-repository generation and auto-research tasks. Muse Spark 1.1 generated challenging coding environments and graded solutions to produce scalable training data for 1.2
• Runtime reliability: A local append-only event log records every model call, tool execution, approval, and edit, providing a single source of truth that is replay-exact and restart-safe—agents resume precisely where they stopped after crashes
• Evaluation benchmarks: Terminal-Bench 2.1 (89 tasks, pass@1 over five attempts), DeepSWE v1.1 (113 tasks across 91 repos and five languages), and Meta Internal Coding Bench (440 tasks from real internal PRs), all executed in isolated Daytona cloud sandboxes with comparisons against Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, and Kimi K3
• Kernel optimization case study: Iterative GPU kernel optimization over 1,000+ tool calls across 24 hours on NVIDIA Hopper GPUs; produced a chunk-parallel + sequential inter-chunk scan for KDA kernels and a two-kernel Triton pipeline reusing shared KV latent for MLA kernels, both outperforming baselines

Industry Insight

The co-training paradigm—where models are trained alongside their operational harness—should become a standard practice for agentic AI systems, as it directly bridges the gap between model capability and tool proficiency; teams building internal coding agents should prioritize similar integrated training approaches rather than treating models and tools as separate concerns. The persistent agent architecture with crash-resilient logging sets a new bar for reliability in long-running autonomous workflows, suggesting that future agent frameworks will increasingly emphasize session continuity and fault tolerance as key differentiators. With Muse Spark 1.2 available via API and the agent in beta, organizations should evaluate integrating these tools into sandboxed review-gated pipelines for repository-scale refactors, test generation, and performance-critical workloads like GPU kernel optimization.

TL;DR

• Meta AI发布Muse Code测试版终端编码代理，由Muse Spark 1.2模型驱动，面向大型代码库的复杂软件工程任务
• 采用持久化异步后台代理架构，代理在整个会话期间保持活跃，避免按任务生成导致的冗余信息收集
• 本地追加事件日志实现重放精确和重启安全，崩溃后可从断点精确恢复，支持长时间运行任务
• Muse Spark 1.2与Muse Code联合训练，覆盖仓库级代码生成、长周期任务和自动研究，具备自我改进能力
• 案例研究显示模型可执行1000+工具调用、持续24小时的GPU内核优化任务，在NVIDIA Hopper上优化KDA和MLA内核

为什么值得看

这篇文章展示了Meta在AI编码代理领域的最新技术突破，为开发者提供了可立即部署的终端编码工具。其架构设计（持久代理、事件日志）和训练方法（联合训练、长周期训练）对构建可靠的AI编码助手具有重要参考价值，同时案例研究验证了模型在复杂工程任务中的实际能力。

技术解析

• 架构设计：Muse Code采用主代理循环加异步后台代理的架构，后台代理在整个会话期间保持活跃状态，而非为单个任务生成。这些专用代理执行下一步操作并自主选择何时向主代理汇报，Meta表示这种持久化设计减少了延迟和对多步骤任务的干预。

• 事件日志机制：Muse Code使用本地追加事件日志作为单一事实来源，记录每次模型调用、工具运行、审批和编辑操作。Meta称之为"重放精确"和"重启安全"，崩溃后代理可精确恢复到停止位置，使长时间运行任务能够容忍故障。

• 训练方法：Muse Spark 1.2与Muse Code联合训练，训练过程包括拒绝采样harness轨迹、目标和子代理的配方优化。训练覆盖整个仓库生成、大型端到端项目和自动研究，模型使用规划、目标条件和上下文压缩来维持进展。Muse Spark 1.1还生成挑战性编码环境和指令遵循模板，通过评分候选方案生成可扩展的训练数据集。

• 基准测试：评估方法包括Terminal-Bench 2.1（89个任务，pass@1 over 5次尝试）、DeepSWE v1.1（113个任务，91个代码库，5种语言）和Meta内部编码基准（440个任务，源自真实内部PR）。对比模型包括Grok 4.5、Claude Opus 5、GPT-5.6 Terra、Gemini 3.6 Flash和Kimi K3，运行在隔离的Daytona云沙箱中。

• 案例研究：GPU内核优化案例中，模型执行1000+工具调用，运行长达24小时。在NVIDIA Hopper GPU上优化KDA和MLA内核，KDA基准为FLA Triton实现（禁止第三方内核库），MLA参考为PyTorch（batch size 1, 64 heads, sequence length 8192, latent dimension 512）。模型为KDA构建了块并行准备内核与顺序块间扫描的组合，为MLA构建了重用共享KV latent作为K和V的双内核Triton管道。

行业启示

• AI编码代理正从单任务执行向长时间、多步骤的复杂软件工程任务演进，持久化架构和容错机制（如事件日志）将成为企业级部署的关键考量因素。
• 模型与工具的联合训练和长周期训练代表了提升AI编码能力的重要方向，未来更多模型将采用类似策略以适配特定工具链和工作流。
• 企业可考虑在沙箱环境中部署此类代理，配合人工审核机制，适用于仓库级重构、迁移、长期bug分类、测试生成和GPU内核优化等场景。