正在打开每日关注
Agent 循环工程：让同一模型持续产出
用 .claude/ 目录、钩子、评估与记忆，把一次性 Agent 变成长程闭环。
来源摘录
THIS IS WHAT LOOP ENGINEERING ACTUALLY LOOKS LIKE INSIDE A .CLAUDE/ FOLDER this folder is why some agents run for hours instead of resetting every turn - not a better model, just what's built around it CLAUDE.md - the contract. operating rules before anything runs settings.json - locks permissions, wires in the hooks hooks/ - the reflexes. pre-tool-use blocks bad calls, post-tool-use logs everything, https://t.co/UC0IBplGsa exits clean agents/ - fresh subagents each time. verifier.md exists just to catch a session lying about what it finished loop-and-harness/ is the part most guides skip. 7 patterns: > turn a short ask into a full spec first > lock "done" into predicates, not vibes > every feature stays strict JSON, starts at passes:false > same 6-step reading ritual every session > no model grades its own work > evaluator gets calibrated so it stops drifting into agreement > strip one component per model release https://t.co/GUQkxBIQq7 fires the loop. write it once, it spawns the next session on its own the repo is public - go check the actual files yourself https://t.co/byjrPpSkAQ
深度解析
是什么 这条推文把“长时间运行的 Agent”归因于 loop engineering：决定效果的不只是更强模型，而是围绕模型建立可重复的规范、钩子、子代理、评估与跨会话记忆。图中以 .claude/ 文件树和 SPEC→LLM LOOP→EVAL 流程，把抽象观点落成工程清单。
核心设计 CLAUDE.md 是运行契约，settings.json 管权限并接入 hooks；pre-tool-use、post-tool-use、stop 分别承担拦截、记录/格式化与干净退出。agents/ 提供新鲜上下文，verifier.md 检查虚假完成。七个模式把短需求扩写成 spec，把 done 写成 predicates，以 strict JSON 和 passes:false 固化状态，用固定六步 session ritual 保持输入质量，再用校准后的 evaluator 抑制自评与迎合；run.sh 触发后续 session，MEMORY.md 负责交接。
关键洞察 它展示的是控制回路设计：把模糊意图变成可检查状态，把一次性回答变成可观察、可迭代的闭环。最值得借鉴的是失败显式化、评估独立性，以及每次模型升级时剥离一个组件来检验 harness 是否仍有必要。
边界与风险 链接正文无法获取，仓库是否真实存在、实现细节和效果数据均未验证，不能把海报当作 README 证据。自动跨会话还会放大错误规范、权限配置与评估偏差；hooks 过多也会增加脆弱性和成本，应配合沙箱、预算、超时、人工审批与真实测试集。