Research agent pattern (safe defaults)

Goal: run a research-style agent that plans, gathers, and stops for your review before any big conclusion.

Best for: competitor scans, “what changed?”, meeting prep—without trusting invented links.

A research agent is not “search forever.” It is a bounded loop: clarify → plan sources → collect with citations → contradict itself → you decide.

Safe default for beginners: browsing/search allowed; sending and file-deletes never; every claim needs a link you can open.

Why beginners skip this

A research agent that concludes without your review ships confident fiction. Safe defaults plan, gather, then stop—claims need sources you opened.

Habit this week (Asia/Shanghai): run one research-style pass with a claim→evidence table; no claim ships without a link you opened—date the table.

Step 1 — Clarify and bound the question

Force a one-sentence question, time range, and audience. Ban a full report until the plan is approved.

You are a research agent (plan only).
Question: [q]
Return: clarified question · 5 sub-questions · source types to prefer · what would falsify a hasty answer · 3 clarifying questions for me
No browsing yet.

Brief: "What changed in EU cookie rules affecting our SaaS checkout in 2026?"
Done for agent: link pack + Unknowns list—NOT a final legal answer.
Stop after gather. I will open primaries and write the conclusion offline.
Ban auto-publish.

Step 2 — Collect with citation rules

After you approve the plan: allow search. Require link + publisher + date. Cap sources (e.g. 8).

Execute research for the clarified question.
Rules: every factual bullet needs link + publisher + date. Max 8 sources. Separate reported vs confirmed.
Output: 1-page brief + contradictions + 3 Monday actions.
As-of: 2026-10-17

Brief: "Summarize three open-source license options for our CLI tool."
Done for agent: link pack to SPDX/official license pages + Unknowns—NOT a pick-for-us verdict.
I choose offline after opening links.

Step 3 — Stress-test before you trust it

Ask the agent to attack its own brief. You open the top links.

Attack this brief: 5 weakest claims and what evidence would change them.
Steel-man the opposite recommendation in 6 bullets.
Do not defend the original brief.

Scene: real work this week (live artifact open).
Attack the Notion vs Claude brief. 5 weakest claims + opposite recommendation. No defending.
Constraints: no invented facts; stop before send/delete; label Unknowns.

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.

Ship the practice

Close the tab only after Practice now is done on a real artifact.

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.

Check the unknown path

What should the model do when a field is missing—null, TBD, or ask you? Write it once in Boundaries.

Red flag

Claims without opened links; treating forum heat as primary; shipping without a counter-question—do not externalize.

Save this card

Card: plan → gather → stop for review · claim→evidence with links you opened · no unopened citations.

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 “run a research-style agent that plans,.” Allow-list tools; leave send/delete off. Guide: agents-research-agent-pattern. 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: run one research-style agent with a written claim→evidence table. No claim ships without a source you opened. Date the table on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed agents-research-agent-pattern, put 25 minutes on the calendar and redo once.

Close the week by quoting aloud the refuse line you actually used once. If you cannot, rerun Practice now and date the pass/fail (Asia/Shanghai).

Pitfalls

Temptation: skip the habit for agents-research-agent-pattern to “save five minutes.”
Job still: run a research-style agent that plans,
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 “Research agent pattern (safe defaults)”—artifact path on Save (Asia/Shanghai).

Today: produce a one-page claim→evidence table from a research-style run; every claim needs a link you opened. Add one counter-question. Date it (Asia/Shanghai).

Today card (copy) — agents-research-agent-pattern: 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

Research agents plan, gather, then stop—claims need sources you opened. This week: one dated claim→evidence table (Asia/Shanghai).

Next

Claim–evidence matrix · Writing pipeline agent · Plan → act → check · Human-in-the-loop checkpoints · Verify before you share

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

调研型智能体模式（安全默认）

目标： 跑一个会计划、采集、并在大结论前停下来等你审的调研智能体。

适合： 竞品扫描、「最近变了什么」、会前准备——且不轻信编造链接。

调研智能体不是「永远搜下去」。它是有边界的循环：澄清 → 计划来源 → 带引用采集 → 自我反驳 → 你来决策。

初学者安全默认：允许浏览/搜索；绝不发送、不删文件；每个主张都要有你能点开的链接。

为什么新手跳过

研究型 agent 未经你审就下结论，等于自信地编。安全默认是：计划、收集、停下——主张必须配你点开过的来源。

本周习惯（Asia/Shanghai）：跑一次带主张→证据表的研究；没打开链接的主张不外传，表格注明日期。

步骤 1——澄清并限定问题

强制一句话问题、时间范围和受众。计划批准前禁止写完整报告。

你是调研智能体（只做计划）。
问题：[问题]
返回：澄清后的问题 · 5 个子问题 · 优先来源类型 · 什么会推翻草率答案 · 问我 3 个澄清问题
先不要浏览。

简报：「2026 年影响我们 SaaS 结账的欧盟 cookie 规则有何变化？」
Agent 的 Done=链接包+未知清单——不是最终法律结论。
收集后停下。我打开一手来源并线下写结论。禁止自动发布。

步骤 2——按引用规则采集

你批准计划后：允许搜索。要求 链接 + 发布方 + 日期。限制来源数量（如 8 个）。

按已澄清问题执行调研。
规则：每条事实要点含 链接+发布方+日期。最多 8 个来源。区分报道与已确认。
输出：一页简报 + 矛盾点 + 下周一 3 件事。
截至：2026-10-17

场景：真实工作（打开活材料）。
简报：「为我们的 CLI 工具总结三种开源许可证选项。」
Done=指向 SPDX/官方许可证页的链接包+未知——不是替我们做选择。
我打开链接后线下决定。
约束：不编造事实；发送/删除前停下；未知处标「未知」。

步骤 3——信任前先加压

让智能体攻击自己的简报。你亲自点开主要链接。

攻击这份简报：5 个最弱主张，以及什么证据会改变它们。
用 6 条要点尽量站在相反建议一边。
不要为原简报辩护。

场景：本周真实活（打开真材料）。
攻击「调研型智能体模式（安全默认）」简报。列出 5 条最弱主张 + 一条相反建议。
不要发明报价。标未知。我打开一手链接前禁止外传结论。

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。

把练习做完

实卡：用「调研型智能体模式（安全默认）」对真实任务跑一条智能体清单——产物路径进 Save（Asia/Shanghai）。

今天把「调研型智能体模式（安全默认）」练完：真实材料、过/不过、一句永不做。注明日期（Asia/Shanghai）。

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。

检查未知路径

字段缺失时模型该怎样——null、TBD，还是问你？在边界里写一次。

红旗

摘要里出现无链接的斩钉截铁结论；把论坛热评或二手转述当一手来源。未做反证一步就外发——立刻停。每条对外主张都要你点开过的来源支撑。

保存这张卡

卡片：计划→收集→停下等人审 · 主张→证据须配你点开的链接 · 未打开的引用删掉。

开始前

开始前先写完成定义（勾选）与人工闸；工具白名单；发送/删除默认关闭。材料是今晚真问题，不是空白框。

本周怎么用

本周限定问题范围，要求每条主张带可点开链接，信任前做一次反证提问。无链接的结论不进对外简报。把过/不过写进笔记并注明日期（Asia/Shanghai）。

常见坑

任务：run a research-style agent that
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

今天（Asia/Shanghai）用研究模式产出一页带引用摘要，并追加一个「如果反面成立会怎样」的反证句。过：每条主张有链接；不过：只有结论段。只收藏标题算失败。

今日卡片（可复制）— agents-research-agent-pattern： 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

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

一句话带走

研究 agent 计划、收集后停下；主张须配你点开的来源。本周交一张带日期的主张→证据表。

下一步

主张–证据矩阵 · 写作流水线 Agent · 计划 → 行动 → 检查 · 人在回路检查点 · 分享前核验

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（research-claim-evidence-matrix）。