Chubby♨️ · @kimmonismus · X·2026-09-01 00:26·1天前
AI 导读

Reflexio 今日正式发布,定位为 AI 智能体的自我改进平台。它能从真实用户对话中提取修正并转化为可复用规则,团队可对照基准响应测试每条规则,效果不佳即弃用。目前客户部署中任务失败率降低 36%,对比响应质量提升 47%,在 GDPval 上相似任务 token 成本下降 57%。

Chubby♨️@kimmonismus
50AI 编辑部评分,满分 100
2026-09-01 00:26· 1天前
AI 导读

Reflexio 今日正式发布,定位为 AI 智能体的自我改进平台。它能从真实用户对话中提取修正并转化为可复用规则,团队可对照基准响应测试每条规则,效果不佳即弃用。目前客户部署中任务失败率降低 36%,对比响应质量提升 47%,在 GDPval 上相似任务 token 成本下降 57%。

AI agents produce a mountain of logs and still repeat the same mistake next week.

Reflexio turns corrections from real user conversations into rules the agent can reuse. Teams can test each rule against a control response and throw it out again if it makes the agent worse.

Yi LuWhat if your agents got better every time they ran? Today we're launching Reflexio — the self-improvement platform for AI agents. It learns from live production...