https://hugozhu.site/tags/skill-distillation/Recent content in Skill-Distillation on All about Raspberry PiHugoenTue, 09 Jun 2026 00:00:00 +0000https://hugozhu.site/post/2026/262-auto-optimize-agent-trajectory/Tue, 09 Jun 2026 00:00:00 +0000https://hugozhu.site/post/2026/262-auto-optimize-agent-trajectory/<p>上个月有人问我一个问题：「我已经有 LLM-as-Judge 做 eval 了，能不能用它来自动优化 Agent 的执行路径？在不降质量的前提下，找到最省钱的轨迹，然后让 Agent 记住？」</p> <p>这个问题的答案值得展开。答案是能，而且这可能是当前 Agent 优化里最值得投入的方向。但大多数团队理解错了「优化」的对象。</p> <p><a href="https://hugozhu.site/img/2026/auto-optimize-agent-trajectory.png"><img src="https://hugozhu.site/img/2026/auto-optimize-agent-trajectory-thumb.jpg" alt="Agent 轨迹优化：从零规划到 Skill 蒸馏"></a></p>