正在打开每日关注
四类 Agent 循环：从手动到全自动
用四种循环解释 Agent 自主度、评估机制与适用场景。
来源摘录
THIS "GETTING STARTED WITH LOOPS" DOC TURNS CLAUDE INTO AN EMPLOYEE THAT SHIPS WHILE YOU SLEEP spoiler: prompt engineering didn't survive the 4 loop types, from you-drive to fully hands-off: > turn-based - 25% autonomy. one task per round, control returns to you. best for exploration when requirements are unclear > goal-based - 70%. you define "done" with /goal and a separate evaluator keeps sending Claude back until it clears the bar. tests passing, target scores, bugs cleared > time-based - 60%. /loop fires on a schedule: check the PR every 5 minutes, address review comments, fix failing CI > proactive - 95%. an event fires, a router dispatches child agents in parallel, a review model oversees the output. no human online at all the advice in it that matters most: loop quality depends on system quality, and tokens are the real cost knowing the 4 types is level one this article is the full 14-step roadmap from typing prompts by hand to designing loops - /goal, SKILL.md, worktrees, MCP, all of it start climbing 👇
深度解析
是什么 这条推文把“让 Agent 持续工作”抽象为四种循环，而不是单次 Prompt：Turn-based 每轮执行一个任务后交还控制；goal-based 由目标和评估器驱动重试；time-based 按时间轮询；proactive 则由事件触发路由，分派子 Agent 并由审查模型收口。
核心设计 四类循环对应触发方式、并发度、完成判定和人类介入频率。配图用 25%、70%、60%、95% 的自主度、人工/Agent 投入条与时间线，把“自主”变成可比较的运行模型；evaluator、trial limit、router、review 分别承担验收、止损、分派和兜底。
关键洞察 选择取决于验收标准和环境稳定性：需求模糊时用 turn-based 探索，结果可测试时用 goal-based，周期性变化适合 time-based，事件量大且任务可拆分时才适合 proactive。循环质量取决于系统约束、上下文、工具权限、观测能力和 Token 成本。
边界与风险 材料没有提供完整 14 步文章、代码或实际评测数据，自主度只是示意。落地仍需处理重试上限、幂等、权限隔离、失败升级、并发冲突与审查偏差；“无人在线”不等于无需人工治理。