Agent Self-Evolution · Part 1

When the Coding Agent Becomes the Optimizer

SkillOpt-Lite does not train model weights. It turns rollout trajectories, a writable skill.md, a coding agent, and a validation gate into a file-based optimization loop—and then stretches the same idea toward harness code.

中文SkillOpt-Lite:当 Coding Agent 成为自己的优化器

从轨迹文件、跨样本诊断和 validation 跑分,读到 HarnessOpt 允许修改的六个文件与当前源码断点。

EnglishSkillOpt-Lite: When the Coding Agent Becomes the Optimizer

Trace trajectory files, cross-sample consensus, validation-gated retention, HarnessOpt permissions, and the gaps visible at HEAD.

SkillOpt-Lite route from a frozen model and harness through skill.md, train trajectories, a coding-agent patch, validation branches, best snapshot, and one-shot test