LlamaFactory作者郑耀威团队开源了PenguinHarness项目，这是一个支持Agent自我进化的开源框架[1]。PenguinHarness基于原创Agent内核，支持GPT-5.6和DeepSeekV4等模型[1]，可以自动完成Agent的构建、评测和持续改进[1]。相比OpenAI Codex，该框架的成本低数十倍，构建一个完整Agent应用仅需0.2元token费用，优化一个Agent从53分提升至95分仅花费0.5元token[1]。
项目已在生产场景落地应用。在医疗领域，体检机构应用该工具后，报告核查时间从30分钟降至几十秒[1]。在制造业，产线停机时间减少65%，产出提升近2倍[1]。团队耗时半年多开发了GDPevo评估基准，覆盖医疗、金融、法律等六大场景[1]。该项目以Apache2.0协议开源，支持Linux、Mac、Windows系统一键安装[1]。
Zheng Yaowei, the creator of LlamaFactory, has released PenguinHarness, an open-source framework that automates agent construction, evaluation, and continuous improvement [1]. The tool achieves a cost reduction of several dozen times compared to OpenAI Codex, with complete agent applications costing as little as 0.2 yuan in token fees [1].
Built on an original agent kernel, PenguinHarness supports models including GPT-5.6 and DeepSeekV4 [1]. The framework enables significant performance gains—one agent improved from a score of 53 to 95 using only 0.5 yuan in token costs [1]. The development team invested over six months creating the GDPevo evaluation benchmark, which spans six major sectors including healthcare, finance, and law [1].
Real-world applications demonstrate substantial operational improvements. A medical examination facility reduced report verification time from 30 minutes to several dozen seconds using the agent system [1]. In manufacturing, companies deploying the technology cut production line downtime by 65 percent while nearly doubling output [1]. PenguinHarness is released under the Apache 2.0 license and supports single-command installation on Linux, Mac, and Windows systems [1].
