腾讯论文提出 SkillAdam，让 AI 自动改进智能体技能指令时记住过往修复、避免大幅重写，从而更稳定高效。在长程购物与旅行规划任务上，它平均准确率达 28.3%，超过此前最佳方法 SkillOpt 的 21.7%，且 token 消耗约为其三分之一。该方法针对自动改写指令常陷入反复、新改动推翻已有修复的问题，为改写 AI 引入两项习惯：记录已修复内容，并在效果不稳定时只做小改动。
New Tencent paper letting AI auto-improve your agent's instructions works better and costs far less if it remembers past fixes and avoids big, risky rewrites.
Agent skills are instruction files that teach an agent how to do a job. Tools that auto-rewrite them often go in circles, burning tokens as new edits undo fixes that already worked.
SkillAdam teaches the rewriting AI 2 habits. It keeps a log of what's been fixed, and it makes smaller changes when results are mixed.
On long shopping and travel planning tasks, it scored 28.3% average accuracy versus 21.7% for SkillOpt, the best earlier method. It also used about a third as many tokens.
If you auto-tune your agent's instructions, give the process a memory of past fixes and a brake on big edits.
– arxiv. org/abs/2609.08944
Title: "SkillAdam: Stable and Efficient Skill Evolution for Agents"
来源：Rohan Paul · x.com