自改进 Agent 接受 harness 修改前应在独立新任务上验证
讲自改进 Agent 怎么防止改坏自己：新改动先在独立任务上跑赢旧版才准上线，还给了 GPT-5.4-nano 在 SpreadsheetBench 拿 77.58% 的例子。
自改进 Agent 会重写固定模型周围的 harness，包括提示词、记忆、工具和代码，被接受的修改会带入后续运行，坏修改可能导致过拟合或丢失旧技能。文章把 harness 自改进方法映射到进化算法，涵盖保留最优搜索、档案库、Pareto 选择和 bandit 引导的预算分配。准入规则是：修改必须在一批新的独立任务上以设定幅度胜过当前版本才能保留。案例中 GPT-5.4-nano 配合进化后的 harness 和技能在 SpreadsheetBench 拿到 77.58%，GPT-5.5 仅做技能调优为 76.20%，但两者成本未对齐。
Self-improving agents should accept a harness edit only after it beats the current version on fresh, independent tasks.
It maps harness self-improvement methods onto evolutionary algorithms, from simple keep-the-best search to archives, Pareto selection, and bandit-guided budgets.
Self-improving agents now rewrite their harness: the prompts, memory, tools, and code around a fixed model. Accepted edits carry into later runs, so a bad edit can overfit, erase old skills, or spread to other users.
They sort these edits by what they change, from prompts to runtime code. Their admission rule: keep an edit only if it beats the current harness on new, independent tasks by a set margin.
In a case they compile from prior work, GPT-5.4-nano with an evolved harness and skills scored 77.58% on SpreadsheetBench. GPT-5.5 with skill tuning alone scored 76.20%, though costs were not matched.
If your agent edits its own setup, test each change on fresh tasks, and keep evaluators, permissions, and emergency stops out of its reach.
