Human-in-the-loop checkpoints

Goal: place simple pause points so you stay in control without micromanaging every keystroke.

Best for: anyone nervous about agents sending, deleting, or spending.

Human-in-the-loop means the agent must stop for your yes/no at risky moments. You are not babysitting every sentence—you are guarding irreversible actions.

Irreversible for beginners: send email/message, delete/move files, buy/subscribe, share outside your device, overwrite production code.

Why beginners skip this

Skipping pauses on send/delete/pay feels smooth until you cannot undo. One rehearsed refusal is cheaper than a recall. Mid-week cost shows up as rework, not as a missing bookmark.

Step 1 — Mark red-line actions

List actions that need a human click. Put them in agent instructions as “STOP and ask.” If your list is empty, you are either doing toy tasks—or you have not admitted what can go wrong.

Common beginner red-lines: send, delete, pay, post publicly, overwrite main. Drafting and summarizing usually are not.

Goal: From this task, list red-line actions that need my approval.
Task: [task]
Format: Red-line list · Why each is risky · Safer alternative (draft-only, sandbox folder)

Task: Draft and send a follow-up to 12 customers who trial-expired last Friday.
List red-line actions, why each is risky, and a safer alternative (e.g. draft-only in Gmail).
Context: Gmail connector is connected; I have never allowed auto-send.

Step 2 — Insert checkpoints into the plan

After the plan exists, add CHECKPOINT lines before each red-line.

Here is the plan:
[paste]
Insert CHECKPOINT: before every red-line action.
Show the revised plan only.

Agent drafts + can send Slack to #customers.
Insert a pause card before send: must show final text + recipients + links; I check tone + no invented SLA; "approve" means I click send myself.
Run once up to the pause only—no send.

Step 3 — Practice a refusal

When the agent asks, you can say no. Teach it: “Paused. Do not retry send. Wait.” Refusal is a skill—practice the sentence once out loud before you need it at 11 p.m.

I refused this action: [action].
Acknowledge pause.
Propose 2 safer next options.
Do not retry the refused action.

Agent can create calendar events and invite externals.
Pause card before invite: show title, attendees, time zone, description; I check no secret URLs; approve = I click invite.
Run up to pause only.

Save this card

Card: red-line actions · pause points · refuse line. Practice one refusal. Keep this beside the tool; refuse to start if a box is blank.

1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)

• [ ] Tools/writes paused
• [ ] Refuse-line dated
• [ ] Artifact path saved
• [ ] Next link chosen (or stop)

Before you start

Write Done (checkboxes) and the human gate on paper before any run for “place simple pause points so you stay in.” Allow-list tools; leave send/delete off. Guide: agents-human-in-the-loop-checkpoints. If you only have a blank practice box, stop and fetch the real job first.

• Goal in one line: ___
• Allowed tools: ___
• Forbidden: ___
• Human approve needed for: ___

How to use this week

This week: insert one pause before send/delete/pay on a live draft. Let the draft auto-run; refuse to skip the human gate. Note which checkpoint you kept on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed agents-human-in-the-loop-checkpoints, put 25 minutes on the calendar and redo once.

Keep it real

Run this guide’s Practice now on work you will actually ship this week—not a fictional scenario. Dated notes beat perfect theory.

Do not bookmark-only

Finish Practice now, then tick this guide on your path. Titles unread are not skills.

Close the loop

After Practice now, write one line: what you will refuse next time. That line is the habit.

Red flag

Red-line actions with no pause; “default approve” as policy; skipped gate with no log—call it an incident.

Pitfalls

Temptation: skip the habit for agents-human-in-the-loop-checkpoints to “save five minutes.”
Job still: place simple pause points so you stay in
Do the smallest correct pass: Done visible · one checkable artifact · Unknowns labeled [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: the artifact + the refuse-line you said aloud.

Practice now

Live card: run one agent checklist item on a real job for “Human-in-the-loop checkpoints”—artifact path on Save (Asia/Shanghai).

Today: pick one agent/job with a send, delete, or pay step. Insert a pause card: what the agent must show · what you check · what "approve" means. Run once up to the pause only.

Today card (copy) — agents-human-in-the-loop-checkpoints: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

• [ ] Artifact path written
• [ ] Refuse example would say "yes"
• [ ] Dated Asia/Shanghai

One sentence takeaway

Pause before send/delete/pay; drafts can auto—irreversible stays human—check it once this week, or it stays a bookmark

• You can point to a dated refuse-line
• You have one artifact path from tonight
• You can teach the red flag in one sentence

Next

Research agent pattern · Define done and permissions · Agent blast radius · Plan → act → check · Failure modes and retries

Pair with today: open the first Next link (agents-research-agent-pattern) only after you dated a refuse-line for this guide (Asia/Shanghai).

人工卡点（Human-in-the-loop）

目标： 在关键处设置简单暂停点，既控得住，又不用盯着每个字。

适合： 担心智能体乱发、乱删、乱花钱的人。

人工卡点的意思是：碰到有风险的时刻，智能体必须停下来等你说可不可以。你不是盯着每句话，而是守住不可逆动作。

对初学者，不可逆包括：发送邮件/消息、删除/移动文件、购买/订阅、把内容分享到设备外、覆盖线上正式代码。

为什么新手跳过

每一步都盯到手软；零检查点又会让草稿直接发出。离开系统前设一个暂停，才是中间路。

步骤 1——标出红线动作

列出必须由人点头的动作。写进智能体说明：「停下来问我。」

目标：根据任务列出需要我批准的红线动作。
任务：[任务]
格式：红线列表 · 为何有风险 · 更安全替代（只起草、沙箱文件夹）

任务：给上周五试用过期的 12 位客户起草跟进，并可发送。
在发送前插入暂停卡：必须展示最终正文+收件人+链接；我查语气与有无编造 SLA；「批准」=我自己点发送。
只跑到暂停——不发送。

步骤 2——把卡点插进计划

计划写好后，在每个红线动作前加一行 CHECKPOINT。

这是计划：
[粘贴]
在每个红线动作前插入 CHECKPOINT：
只展示修订后的计划。

场景：真实工作（打开活材料）。
Agent 可起草并可往 #customers 发 Slack。
发送前暂停卡：展示最终文本+频道+提及；我查事实；批准=我点发送。
只跑到暂停。
约束：不编造事实；发送/删除前停下；未知处标「未知」。

步骤 3——练习一次拒绝

智能体来问时，你可以否。教它：「已暂停。不要重试发送。等待。」

我拒绝了这个动作：[动作]。
确认暂停。
给出 2 个更安全的下一步选项。
不要重试被拒动作。

场景：真实工作（打开活材料）。
Agent 可建日历事件并邀请外部人。
邀请前暂停卡：展示标题、出席者、时区、描述；我查有无秘密 URL；批准=我点邀请。
只跑到暂停。
约束：不编造事实；发送/删除前停下；未知处标「未知」。

保存这张卡

卡面：红线动作列表 + 卡点位置 + 拒绝话术。跑前看卡。开跑前默念一次。

1. 我交出了：___ 2. 我拒绝了：___ 3. 我永不再做：___（日期：Asia/Shanghai）

• [ ] 工具/写入已暂停
• [ ] 拒做句已标注日期
• [ ] 成品路径已存
• [ ] 已选下一步（或停）

开始前

打开你真正会用的账号/项目。准备好与「人工卡点（Human-in-the-loop）」相关的真实材料（不要空白练习框）。

默认姿态：在人确认发送/写入之前只出草稿。授予新工具前，先写一条带日期的拒做句（Asia/Shanghai）。

• 一句话目标：___
• 允许的工具：___
• 禁止：___
• 需要人批准：___

本周怎么用

本周标出红线动作，把人工卡点插进计划。演练一次「拒绝继续」并写入日志，证明门闸真的会停。把过/不过写进笔记并注明日期（Asia/Shanghai）。

保持真实

用本周真会交付的工作做「现在就练一次」——不要虚构场景。带日期的笔记胜过完美理论。

不要只收藏

做完「现在就练一次」，再在路径上打勾。没练过的标题不是技能。

闭环

「现在就练一次」后写一行：下次你要拒绝什么。那一行才是习惯。

红旗

红线动作无卡点；把「默认同意」写成策略。卡点被跳过却无日志——当作事故。卡上写不清就不继续——先补检查项再跑。

常见坑

任务：place simple pause points so you
场景要点：Agent 运行前 Done/门闸写好；练习编号 4/4。
边界：仅草稿；未知 [?]；禁止悄悄发送；禁止编造指标。
返回：可检查产出 + 过/不过理由一行。

现在就练一次

实卡：用「人工卡点（Human-in-the-loop）」对真实任务跑一条智能体清单——产物路径进 Save（Asia/Shanghai）。

今天在真实计划里插入至少两处卡点，并口头拒绝一次越权请求。过/不过与永不做注明日期（Asia/Shanghai）。

今日卡片（可复制）— agents-human-in-the-loop-checkpoints： 1. 我要产出的成品：___ 2. 通过/失败标准：___ 3. 若想跳过核验/发送的拒做句：___（日期：Asia/Shanghai） 4. 本篇一条永不做：___

• [ ] 成品路径已写
• [ ] 拒做例句会说「可以」
• [ ] 已标注 Asia/Shanghai 日期

一句话带走

发送、删除、支付前必须暂停等人拍板；草稿可以自动跑，不可逆动作绝不跳过人工卡点。本周留下一处你坚持保留的卡点记录。本周做一次可检查练习，否则只是收藏。本周留下可核对的过/不过记录并注明日期（Asia/Shanghai）。

• 能指出一条带日期的拒做句
• 今晚有一条成品路径
• 能用一句话说清红旗

下一步

研究型 Agent 模式 · 定义完成与权限 · Agent 爆炸半径 · 计划 → 行动 → 检查 · 失败模式与重试

今日搭配：先为本篇写上拒做句日期（Asia/Shanghai），再打开下一步里的第一篇（agents-research-agent-pattern）。