Products

新加坡 Agnes AI 用自我复盘循环让王者荣耀胜率从 0% 提到 60%

Why we picked it

做智能体的都知道,agent 反复犯同一个错、最后还是得手动改提示词有多烦。这个新加坡团队搞了个 coach 模型自己复盘《王者荣耀》对局、重写 playbook、测试通过才保留,三轮就把胜率从 0 打到 60%,还不用重训模型。写 agent 工作流的人值得看看这套闭环怎么搭的。

新加坡 AI 实验室 @agnesai_sapiens 构建了一套自学习系统,让智能体战队在《王者荣耀》中的胜率从 0% 提升到 60%。前 192 局全部落败,经过三轮迭代后胜负比转正,全程没有对模型做任何重训练。

Body

机制是一个闭环:coach 模型复盘已结束的对局,找出问题并重写战队的 playbook;每项改动必须通过测试才能保留,否则被丢弃。也就是说,改进全部发生在策略与提示层面,而不是通过梯度更新。

这套系统的价值在于回应了当前智能体开发的普遍痛点:智能体往往反复犯同一个错误,直到有人手动修改提示词为止。自动审查自身表现、只保留有效改动的循环,为智能体自我改进提供了一条不依赖重训练的工程路径。

Original · kimmonismus

A Singapore AI lab built a self-learning system that went from 0% to 60% win rate in Honor of Kings. It took three rounds, and no model was retrained.

@agnesai_sapiens built a loop where a coach model looks at the finished games, figures out what went wrong and rewrites the team's playbook. A change only stays if it passes the tests.

What's impressive is how fast it learned. The team won none of its first 192 games, and three rounds later it was winning more than it lost.

Honestly, the most annoying thing about agents right now is that they make the same mistake over and over until someone fixes the prompt by hand. A system that reviews its own work and only keeps what actually works is exactly what I want to see more of.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Products" desk, focusing on Southeast Asia and the ASEAN-6 (Singapore, Malaysia, Indonesia, Thailand, Vietnam, Philippines) digital economy. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Agnes AI, agents, automation, recursive self-improvement.

Source attribution: curated from kimmonismus (@kimmonismus).