https://chenyang-zheng.github.io/tags/long-running-agents/Recent content in Long-Running-Agents on Chenyang's EurekaHugoen-usTue, 09 Jun 2026 13:38:00 +0800https://chenyang-zheng.github.io/posts/adversarial-agile-loop/Mon, 01 Jun 2026 20:02:20 +0800https://chenyang-zheng.github.io/posts/adversarial-agile-loop/长时Agent常陷入低效循环——不断收集失败却无法沉淀出可复用的规则。本文在Anthropic的Planner–Generator–Evaluator架构和Ralph Loop基础上，引入敏捷开发的user story等作为最小可观测航点，通过对抗式评估和语义残差回传，把失败转化为下一轮的显式约束，提高Agent从失败中学习的效率。