Anthropic 工程师公开智能体记忆与工作流图设计方法
Anthropic 公开内部智能体记忆与工作流图方法，核心是让 AI 改进可累积，值得从业者借鉴。
Anthropic 工程师 Lamis 在伦敦 AI Native DevCon 上拆解内部方法：用文件系统作为 agent 记忆，支持回滚、防冲突与权限管理；通过“做梦”机制让独立 agent 复盘全局，找出反复出错之处；并用假边测试优化工作流图，将串行步骤改为并行钻石结构，让 AI 的改进可累积。
正文摘录
 Reported by Synced  "Stop writing prompts for Claude. Build it a graph it can run on its own." Recently, Anthropic engineer Lamis took the stage at London's AI Native DevCon and spent nearly 30 minutes publicly deconstructing an internal methodology for making AI "smarter the more you use it." She opened with a joke: "You've been listening to talks all day — your 'context window' must be nearly full by now. Hopefully you still have a bit of headroom, because what I'm about to talk about is exactly that: context." After the laughter, she proceeded to spill a year's worth of Anthropic's secrets. ![](https://aiera.com.cn/wp-content/uploads/20…