Sentient 发布 EvoSkill v2，让 AI 代理自我改进无需重新训练

在修复电子表格任务中，该技术使最难的 21 个任务从 120 次尝试中 3 次成功提升至 21 次成功。该技术用于测试 OpenAI-Hugging Face 事件中代理试图攻击自身评分器的场景。

Agents can self-improve without retraining. EvoSkill v2 achieves this with persistent agent skills. A coach agent reads the failed runs and writes the skill. The worker loads it the next time a similar task shows up. No weights are touched. Every improvement comes from a simple file with lessons. Every bad lesson also gets saved. On spreadsheet repair, the coach found the grader trusted cached values and wrote a skill telling the worker to skip recalculation. Sentient's fix was to split the roles. The agent that writes skills cannot touch the test. A person reviews the results after every round. With that in place, the hardest spreadsheet tasks went from 3 passes out of 120 to 21. Sentient @SentientAGI Last week, Dario Amodei published "We Must Pace the Frontier". His concern: the OpenAI–Hugging Face incident in which a swarm of agents tried to hack their own grader. Rather than take his word for it, we used EvoSkill to test it by building a coach whose job was to make another AI score higher on a test. Here’s what happened ↓ 🔗 View Quoted Tweet📊 10 ⚡

本条情报归属 AITOP「模型」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「中美大模型技术代差与商业化对照」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：Sentient、EvoSkill、代理、自我改进，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明：本文原始线索采编自「elvis」（@omarsar0），经 AITOP 多维评分管线结构化提炼。截至目前，全网已有 11 家独立信源对该事件进行交叉印证与跟进报道（含 Gary Marcus、VnExpress Số hóa、Rappler: Technology 等），属于近期高关注度行业信号。