Prime Intellect团队推出了Prime Agent，这是一套基于递归语言模型（RLM）和持续强化学习框架的自我改进编码智能体系统[1]。该系统采用IPython REPL和持续强化学习设计，使智能体能够动态管理自身的提示词、技能、记忆和子智能体[1]。
Prime Agent在ARC-AGI 3基准测试中取得了突出成绩，RHAE Best@1达到95.5%，超越了95.4%的人类专家基准[1]。在Best@3指标上，该系统获得99.97%的成绩，完成了全部183关卡[1]。相比各类模型原生框架，Prime Agent在更低的token使用量下实现了更高的分数[1]。
该系统已完全开源发布[1]。Prime Agent支持持久化子智能体、跨session通信和自主模式运行[1]。系统还引入了自我改进管道，能够基于智能体的运行轨迹自动优化系统提示词和技能[1]。
The Prime Intellect team has unveiled Prime Agent, an open-source artificial intelligence system designed to autonomously improve its own capabilities through continuous learning and self-refinement [1]. Built on recursive language models (RLM) and a continuous reinforcement learning framework, the system employs an IPython REPL environment that enables agents to dynamically manage their own prompts, skills, memory, and sub-agents [1].
In benchmark testing, Prime Agent demonstrated exceptional performance on the ARC-AGI 3 challenge, achieving a 95.5% success rate on the RHAE Best@1 metric, surpassing the 95.4% human expert baseline [1]. The system completed all 183 levels of the benchmark with a 99.97% Best@3 score [1]. Notably, Prime Agent accomplished these results while consuming fewer tokens than comparative frameworks running on native model implementations [1].
The system introduces a self-improvement pipeline that automatically optimizes its system prompts and skills based on agent trajectories [1]. Additional features include support for persistent sub-agents, cross-session communication, and an autonomous operation mode [1]. Prime Agent is fully open-source [1].