Raven：为递归自我改进设计的多智能体编排框架

EverMind AI 推出了 Raven，一个专为递归自我改进而构建的多智能体编排框架1。该系统采用主控智能体架构，集成了四个专用智能体：Raven-Research 负责研究、Raven-Code 负责编码、Raven-Design 负责设计，以及 Raven-Oncall 用于自动化工作流1。框架内部采用模块化设计，包含 Memory、Planning、Capability 和 Action 四个解耦的策略模块1。

Raven 能够自主规划、执行和评估复杂项目。在实际演示中，该系统在 4 天内自主完成了 42 轮规划、开发和验证周期，成功构建了一款 Godot 4 第一人称射击游戏1。在 AI 自我改进的案例中，Raven 在 nanochat 预训练实验里完成了 172 次训练运行，在 20 分钟、单 GPU 预算的约束下将验证困惑度（val_bpb）降低了 5.8%1。此外，该框架支持通过 ACP、CLI 或 OpenAI 兼容 API 连接 13 个第三方智能体1。

EverMind AI has introduced Raven, a multi-agent orchestration framework designed to enable recursive self-improvement through coordinated autonomous agents.1 The system functions as a master controller that integrates four specialized agents—Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall—each dedicated to specific tasks within complex project execution.1 Through this architecture, Raven can autonomously plan, execute, and evaluate projects.1

The framework has demonstrated its capabilities through several completed project implementations.1 In one instance, Raven independently completed a Godot 4 first-person shooter game within four days, cycling through 42 rounds of planning, development, and validation.1 Additionally, the system achieved measurable improvements in AI model optimization during nanochat pretraining experiments, executing 172 training runs and reducing validation bits-per-byte (val_bpb) by 5.8% within a 20-minute window using a single GPU budget.1

Raven supports integration with 13 third-party agents via ACP, CLI, or OpenAI-compatible APIs, enabling extended functionality beyond its core components.1 The framework employs a modular architecture comprising four decoupled strategy modules—Memory, Planning, Capability, and Action—designed to provide flexibility and scalability.1