I think the Agent should be able to learn from experience. I’ve worked on a related project—please take a look to see if it might be helpful for the harness.

[LittleRockets]

AgentExperience is a framework-neutral instrumentation and memory library designed to give AI agents "memory for what actually worked." Unlike traditional memory systems that store raw conversation histories or model-generated summaries, AgentExperience captures baseline-relative strategy deltas backed by deterministic evidence. If it makes sense and helps with the harness, that would be something to be happy about. https://github.com/LittleRockets/AgentExperience.git

[shinjiyu]

官方创造模式已经能在内存里试 Cordis 插件，重启会清掉。我们补的是试验之后：信号 → 打分 → 固化到 .evolve/ 或回滚。不是「已经自动自进化」，也不改 agent-loop。

dsh plugin --profile web add github:shinjiyu/deepseek-harness-evolver

介绍：#1720