Research · Deep Research

1. Introduction

1.1 Background

Deep research is one of the first directions in which agents achieved productization at scale. Since late 2024, "Deep Research" has rapidly evolved from a product feature name into an independent capability category: given a question, automatically complete retrieval, reading, cross-validation, synthesis, and writing.

Its productization pace has been extremely fast: 2024-12 Google launched Gemini Deep Research for consumers; 2025-02-02 OpenAI released Deep Research in ChatGPT, based on a variant of o3 trained with end-to-end reinforcement learning for browsing and reasoning tasks; 2025-02-15 Perplexity opened it free to all users; 2025-06-26 OpenAI turned Deep Research into a code-callable model via the Responses API; 2025-07-16 LangChain open-sourced a reference implementation of the three-stage pipeline; 2026-02-10 OpenAI added MCP and connector support and allowed users to restrict the search scope to trusted sites.

This evolutionary path itself is a microcosm of Harness's maturation: from "a model that can search" to "a research runtime that can be orchestrated, source-restricted, and integrated with external tools".

At the same time, the industry baseline for open science and reproducibility is also rising in parallel. According to a ten-year survey cited by Springer Nature's official blog: about 80% of researchers worldwide have heard of the FAIR principles, and the proportion who have "never heard" of them fell from about 60% to 20%; the survey has cumulatively covered 212 countries/regions and more than 43,000 researchers since 2016. One respondent put it this way: "Open data has gone from novelty to mainstream." This provides the industry context for deep research's citation verifiability requirement.

1.2 Definition

An AI Harness for deep research refers to the engineering carrier layer that, around literature retrieval, evidence synthesis, reviews, and the automation of research reports, provides agents with source governance, retrieval orchestration, evidence-table management, citation verifiability, and cost control.

Its deliverable is not "a fluent article" but an evidence set in which every conclusion can be traced back to a verifiable source.

Boundary clarification:

• A deep research Harness is not an "automatic report-writing robot". Fluent but unverifiable text is considered a failed deliverable in this direction.
• The key difference from data analytics (02-analytics) lies in ground truth: data analytics has an executable result set, deep research does not. This difference determines that the forms of their L5 evaluation layers are entirely different.

1.3 Position in the AI Harness System

图 1-1｜深度研究 Harness 六层定位（L1–L6）

数据来源：基于本文分析绘制的示意图。

The core lies in L1 and L3 (assigned per task), but the real bottleneck is L5.

L1 and L3 are the core:

• L1: one of the essential values of multi-agent systems is context isolation — each sub-agent has an independent context window, preventing a single window from being polluted by a flood of retrieval results. This is the prerequisite for research tasks scaling to hundreds of sources.
• L3: the scope / research / write three-stage pipeline has become a de facto standard; orchestrator-worker is the proven winning architecture.

L5 is the most prominent shortcoming: this direction lacks ground truth. A review has no standard answer, only "whether it is verifiable, whether it covers the main evidence, and whether there is systematic bias". Feasible alternatives include: citation verifiability rate, external benchmarks (e.g., DRACO, SealQA-like), and a hard DoD that "every conclusion must have a clickable source".

It is worth calling out: this shortcoming makes this direction one of the highest hallucination-amplification risk directions. If the data analytics direction computes something wrong, there is a result-set comparison to catch it; in the deep research direction, a fabricated-looking citation, without a line-by-line verification mechanism, can travel all the way into the final report.

1.4 Value and Bottlenecks

Value:

• Compresses hours into tens of minutes. A common description: a deep research agent runs for 5–30 minutes and produces results that a human analyst would need half a day to complete; one vendor puts it as "completing in tens of minutes what would take humans hours".
• Retrieval breadth surpasses humans. Dozens of searches, hundreds of sources, and multiple rounds of refinement are hard for a human to complete in the same time.
• Sources can be restricted. Since 2026-02-10, users can restrict the search scope to trusted sites — this is essentially the Harness-layer source-allowlist capability being productized.

Bottlenecks:

• No ground truth: no automated metric can judge whether a review is "right".
• Hallucinated citations: model-generated citations may be complete in format but point to literature that does not exist or does not support the claim.
• Cost and time: a single deep research session often costs $10–30 (based on mid-2026 pricing, which changes frequently); runtimes of 5–30 minutes limit the interactive experience.
• Evidence bias: retrieval ranking and the source allowlist systematically affect conclusions; without disclosure, readers have no way to judge.

2. Glossary

3. Case Studies

3.1 Case 1: The Productization Chronicle of Deep Research

3.1.1 Background

Deep research completed its transformation from a product feature to a capability category within two years. Understanding this chronicle helps judge which capabilities have already been commoditized (usable directly) and which still need to be built by Harness itself.

3.1.2 Approach

From a Harness perspective, this chronicle contains three key architectural leaps:

1. Callability (2025-06-26): Deep Research went from "a button in a product" to "a model in an API". This step made it possible to embed research capability into one's own Harness rather than use it only within a vendor's interface.
2. Open-source reference implementation (2025-07-16): the three-stage scope / research / write pipeline was open-sourced, and the model, search tools, and MCP servers are all pluggable. This gave self-built Harnesses a directly reusable orchestration skeleton.
3. Source restriction and tool integration (2026-02-10): supports MCP connectors and allows restricting the search scope to trusted sites. This step built L6 governance capability (source allowlist) directly into the product.

3.1.3 Results

The capability-level improvement is significant: HLE rose from 26.6% in 2025-02 to 46.4% in 2025-12 (mentioned by multiple sources, fairly high confidence, but it is advisable to verify against official releases); on GAIA, open-source solutions rose from 46% to 55.15%, compressing the gap between open and closed source into a catchable range.

As for runtime characteristics: Perplexity Deep Research produces a report in about 2–4 minutes, involving dozens of searches, hundreds of sources, and 3–5 refinement rounds; ChatGPT Deep Research takes about 5–30 minutes and processes hundreds of pages.

As for cost (mid-2026 pricing, which changes frequently): the OpenAI o3-deep-research API is $10 per million input tokens and $40 per million output tokens, and a single deep research session often costs $10–30; ChatGPT Pro is $200/month; Google AI Pro is $19.99/month; Linkup's /research is $0.25–2.50 per request.

The lesson for self-built Harnesses is: the capability layer is already highly commoditized, and the room for differentiation lies in the governance and evaluation layers — restricting sources, verifying citations, controlling cost, and curating evidence tables are things the product will not do for you.

3.2 Case 2: Multi-Agent Architecture and Context Isolation

3.2.1 Background

A single agent doing deep research faces a structural contradiction: research needs to cover hundreds of sources, but a single context window cannot hold that much information; force-filling it causes context pollution, and later-stage reasoning gets skewed by earlier retrieval results.

3.2.2 Approach

Anthropic's Claude Research engineering blog, published 2025-06-13, disclosed the architecture of its multi-agent research system:

• orchestrator-worker architecture: one orchestrating agent decomposes the task, dispatches multiple sub-agents to retrieve in parallel, and then aggregates the results.
• Context isolation: each sub-agent has an independent context window and handles only its own sub-question, preventing a single window from being polluted.
• Token amplification: the system's disclosed token usage is roughly 15× that of ordinary chat.
• Evaluation method: it adopts end-state evaluation — it does not judge the process step by step but only evaluates the final deliverable.

Anthropic's internal evaluation states: a multi-agent research system with Claude Opus 4 as lead and Claude Sonnet 4 as subagents was "90.2% ahead" of a single Claude Opus 4 agent on the same task. The original phrasing is semantically ambiguous (likely a relative improvement rather than an absolute difference), marked [To be verified], and should be cited only after checking the original Anthropic engineering blog.

Another independent piece of evidence comes from Google's benchmark performance: Gemini Deep Research reached 94.3% on GPQA Diamond (Gemini 3.1 Pro base); on BrowseComp with full tool access, Gemini 3.1 Pro reached 85.9%, a 45% improvement over the previous generation. Tool access itself may contribute more to results than model generational differences — consistent with Harness's claim that "tools and context determine the ceiling".

3.2.3 Results

This case yields three transferable Harness design conclusions:

1. Context isolation is better than context compression. Rather than laboriously compressing hundreds of sources into a single window, distribute them across multiple independent windows and compress only at the aggregation layer.
2. End-state evaluation suits research tasks better. The research process is hard to score step by step, but whether the final deliverable "answers the question and whether every conclusion is verifiable" is decidable.
3. A 15× token amplification is the norm and must be included in the cost model. The cost of multi-agent research cannot be estimated as a single conversation; Harness must provide budget guardrails at L6.

3.3 Case 3: How to Evaluate When There Is No Ground Truth

3.3.1 Background

Deep research is the direction with the weakest ground truth in this group. A review has no standard answer: neither an executable criterion like "whether the result sets match" nor a "correct answer" can be defined.

This predicament shows up directly in commercial products as evaluation opacity: reportedly, only Linkup has published a reproducible accuracy benchmark (61% on SealQA-0); Valyu leads on DRACO (DeepResearch-Bench) with 53.1; and OpenAI and Gemini have both refrained from publishing independent reproducible accuracy benchmarks for their research APIs.

Partial comparison of the three vendors on public benchmarks (as of 2025-12):

3.3.2 Approach

In the absence of ground truth, the feasible alternative evaluation approaches have four tiers:

Among these, citation verifiability should be a hard DoD: every conclusion must have a clickable source, and the source's original text must support the claim. This rule looks simple, yet it intercepts this direction's most common failure — hallucinated citations that are complete in format but point to the wrong thing or do not support the claim.

3.3.3 Results

Citation verifiability rate can become an effective substitute metric because it converts an "undecidable correctness question" into a "decidable provenance question": we cannot automatically judge whether a review's conclusions are right, but we can automatically judge whether each of its citations actually exists and supports the corresponding claim.

The cost of this conversion is that it can only guarantee that sources are real, not that inferences are valid. Hence this direction requires conclusions to be graded into three tiers — "Verified / Partially verified / Not verified" — where only "citation verifiable and inference directly derived from the original text" counts as Verified, and "citation verifiable but requiring cross-source inference" counts as Partially verified.

It should be clarified: the evaluation data involved in this case (SealQA-0 61%, DRACO 53.1) contains vendor self-reported components, is marked [To be verified], and is not an independent basis for capability comparison.

4. Practice Standards

4.1 AGENTS.md Specification

Standard source statement: the following is the proposed standard draft of the deep research direction's AGENTS.md of this article. To date, there is no official or industry-recognized standard original text; AGENTS.md is a community convention rather than a standard. This draft inherits all clauses of the data science group-level AGENTS.md.

# AGENTS.md —— 深度研究（Research / Deep Research）

> 继承数据科学组级 AGENTS.md 全部条款。本节为深度研究方向的收紧与扩展。
> 本文件为标准建议稿，业界尚无官方标准版本。

## 角色与边界

- 本 Agent 是**证据的搜集与综合者**，不是观点权威、不是事实来源。
- 可以做：检索、阅读、摘录、交叉验证、生成证据表、按证据成文、标注不确定性。
- 不可以做：生成未经检索的引用、把模型记忆当作事实、在证据不足时给出确定结论、代表组织对外发布观点。
- **引用只能来自检索结果，禁止由模型凭记忆生成。** 这一条没有例外。

## 环境假设

- 存在搜索与浏览工具（可带来源白名单），且返回内容附带可点击 URL。
- 存在文献数据库 / 内部知识库接入（如 MCP 连接器）。
- 存在引用管理结构：可记录 URL、标题、作者、发布时间、访问时间、原文片段。
- 存在成本可见性与预算上限。
- 研究范围（子问题清单）可被显式表达并随进度更新。

## 上下文加载顺序（Context Budget）

1. 研究问题与范围声明（必须）
2. 已确认的子问题清单与优先级（必须）
3. 已收集的证据表（去重后，按相关性排序）
4. 权威来源的原文片段（优先一手来源）
5. 冲突证据与反方观点（**必须显式保留，不得裁剪**）
6. 术语表与领域背景
7. 历史研究计划与既往结论
8. 用户偏好（详略、格式、语言）

- **上下文隔离优先于上下文压缩**：子问题分给独立子智能体处理，只在汇总层压缩。
- 冲突证据不得因「与主结论不符」而被剔除。

## 工具契约

| 工具 | 用途 | 模式 | 约束 |
|---|---|---|---|
| 搜索工具 | 检索公开信息 | 只读 | 支持来源白名单；记录查询词与返回条数 |
| 浏览 / 抓取 | 读取具体页面 | 只读 | 记录 URL 与抓取时间；超长页面分段读取 |
| 文献数据库 | 检索论文与预印本 | 只读 | 记录检索式与命中数 |
| 内部知识库 | 检索组织内部资料 | 只读 | 遵守内部分级与权限 |
| 引用管理 | 写入引用条目 | 受限写 | 每条引用必须含 URL 与原文片段 |
| MCP 连接器 | 接入外部数据源与工具 | 受限 | 仅启用已登记的连接器 |

- 工具调用记录：查询词、返回条数、被采纳条数、耗时、成本。
- **禁止使用来源白名单之外的站点**，除非用户显式授权并留痕。

## 数据源契约

- 每条证据必须记录：URL、标题、作者或机构、发布时间、访问时间、原文片段。
- 优先一手来源（原始论文、官方文件、一手报道）；二手转述必须标注为转述。
- 证据必须标注可信度等级：一手 / 权威二手 / 一般二手 / 存疑。
- **缺失的字段必须显式写「未知」，禁止留空或编造**。
- 付费墙或不可访问来源不得被当作已核验来源。

## 任务执行流程（SOP）

1. **Scope（界定范围）**
   复述问题；拆解为子问题清单；确定检索策略、来源白名单、预算与时间上限；列出待确认假设。
2. **计划卡点**
   范围与预算经人确认后进入检索阶段。
3. **Research（检索与阅读）**
   按子问题并行检索；每轮记录查询词与命中；优先抓取一手来源；持续更新证据表。
4. **交叉验证**
   对关键结论寻找至少两个独立来源；记录冲突；不因与主结论不符而删除证据。
5. **缺口识别**
   识别未覆盖的子问题与证据不足的结论；补充检索或标记为未验证。
6. **Write（综合成文）**
   按证据表成文；每条结论附来源编号；区分事实、推论与判断。
7. **引用核验**
   逐条打开引用，确认 URL 可达且原文片段支持对应论断；核验失败的引用必须删除或降级。
8. **分级与交付**
   结论分「已验证 / 部分验证 / 未验证」三档；输出证据表、不确定性与建议。
9. **归档**
   保存研究计划、证据表、引用库与成本记录，使研究可被复现与更新。

## 验证与证据要求

- **每条结论必须有可点击来源**，且来源原文支持该论断。
- **引用可核查率**作为硬性指标：核验失败的引用必须删除或降级，不得保留以充数。
- 关键结论需至少两个独立来源支持；单一来源的须标注。
- 必须主动呈现**反方观点与不利证据**，并设置独立章节。
- 必须区分「原文事实」「跨源推论」「模型判断」三类表述。
- 结论分「已验证 / 部分验证 / 未验证」三档列出，未验证结论只能列为假设。
- 数值、日期、人名、机构名必须与来源一致，禁止「顺手修正」。

## 失败与升级策略

- 同类检索无果重试不超过 2 次；第 3 次改变检索式或升级。
- **引用核验失败**：删除或降级该引用；若因此导致结论失去支撑，把结论降级为未验证。
- **来源不可访问**：标注为不可核验，不得当作已核验来源。
- **证据冲突**：并列呈现冲突双方与可能的解释，不得自行择一。
- **成本超预算**：暂停检索，报告已发生成本与已覆盖范围，申请追加或缩减范围。
- **领域超出能力**：明确说明缺乏领域判断力，建议人工专家复核。
- 升级时携带：研究问题、子问题覆盖情况、证据表、冲突点、成本、建议下一步。

## 安全与合规红线

- 不得生成未经检索的引用、文献、标准编号或链接。
- 不得把模型记忆当作事实来源。
- 不得删除或隐藏与主结论冲突的证据。
- 不得绕过来源白名单。
- 不得在证据不足时给出确定性结论。
- 不得侵犯版权：引用须注明出处，长篇原文摘录须限制篇幅并标注。
- 不得在未标注的情况下把他人观点表述为自己的分析结论。
- 涉及个人信息、未公开信息与受限内部资料的检索，须确认权限后方可进行。

## 禁止事项

- 禁止凭记忆写引用。
- 禁止用「有研究表明」「据统计」等无来源表述。
- 禁止只呈现支持性证据。
- 禁止把推论写成事实。
- 禁止在未核验的情况下保留引用以充数。
- 禁止用检索条数代替证据质量。
- 禁止跨方向复制通用模板；深度研究的引用核验要求与科学计算、HPC 有实质差异。

## 输出格式

- 结论先行 → 证据表 → 反方观点与冲突 → 不确定性 → 建议。
- 证据表列：结论、来源编号、URL、来源类型、可信度等级、原文片段、核验状态。
- 引用使用编号，正文标注如「见来源 [3]」。
- 数值带单位与来源；范围用「~」连接；百分比数值与 % 之间无空格。
- 每个输出必须包含：覆盖范围（检索了什么、没检索什么）、成本记录、核验状态统计。

## 评估与自检

- [ ] 研究范围与子问题清单已确认
- [ ] 每条引用都来自检索，无一由模型生成
- [ ] 引用可核查率已统计，核验失败的已删除或降级
- [ ] 关键结论有至少两个独立来源
- [ ] 反方观点与冲突证据已独立呈现
- [ ] 事实、推论、判断已区分标注
- [ ] 结论已分三档（已验证 / 部分验证 / 未验证）
- [ ] 覆盖范围与未覆盖的子问题已说明
- [ ] 成本在预算内，或已上报
- [ ] 研究计划、证据表与引用库已归档

4.2 SKILL.md Specification

Standard source statement: the following is the proposed standard draft of the deep research direction's SKILL.md of this article, for which likewise no official standard original text exists.

---
name: deep-research-report
description: 深度研究与调研报告的标准执行流程与验收基线。适用于文献综述、竞品与市场调研、技术选型论证、政策与标准梳理等需要可核查证据的任务。触发场景：任何需要产出带引用的研究结论的任务。
version: 1.0
created: 2026-09-12
---

# 深度研究标准流程

## 适用场景

- 文献综述与系统性梳理。
- 技术选型、竞品与市场调研。
- 政策、标准与法规梳理。
- 需要证据支撑的内部论证材料。
- 不适用场景：纯观点性写作、不依赖外部证据的创意内容、需要原创实验数据回答的问题。

## 前置条件

- 研究问题可拆解为有限的子问题清单。
- 搜索与浏览工具可用，且支持来源白名单。
- 引用管理结构可用（URL、标题、作者、发布时间、访问时间、原文片段）。
- 预算与时间上限已设定。
- 输出语言、详略与格式要求已明确。

## 输入

| 输入项 | 必需 | 说明 |
|---|---|---|
| 研究问题 | 是 | 需回答的核心问题 |
| 范围与边界 | 是 | 时间范围、地域范围、纳入与排除标准 |
| 来源白名单 | 否 | 限定可检索的站点或数据库 |
| 预算与时间上限 | 是 | 成本上限、最长运行时间 |
| 输出格式要求 | 是 | 篇幅、结构、语言 |
| 既有资料 | 否 | 已知的种子文献或内部资料 |

## 输出

| 输出项 | 必需 | 说明 |
|---|---|---|
| 结论摘要 | 是 | 直接回答研究问题，分三档标注验证状态 |
| 证据表 | 是 | 结论 × 来源 × 可信度 × 核验状态 |
| 报告正文 | 是 | 每条结论带来源编号 |
| 反方观点与冲突 | 是 | 独立章节，不得省略 |
| 覆盖范围说明 | 是 | 检索了什么、没检索什么、为何 |
| 引用库 | 是 | 全部引用的完整元数据与访问时间 |
| 成本记录 | 是 | token 或请求数、耗时、费用 |

## 执行步骤

1. **界定范围（Scope）**
   复述问题并拆解子问题；定义纳入排除标准；确定来源白名单、预算与时间上限；输出研究计划。

2. **计划卡点确认**
   研究计划经人确认后进入检索；未确认不启动大规模检索。

3. **并行检索（Research）**
   按子问题分派检索；记录每轮查询词、命中数与采纳数；优先一手来源；持续更新证据表。

4. **阅读与摘录**
   抓取并阅读关键来源；摘录支持结论的原文片段，记录 URL 与访问时间；超长来源分段处理。

5. **交叉验证**
   对每条关键结论寻找至少两个独立来源；记录一致与冲突；冲突证据保留。

6. **缺口识别与补检**
   列出未覆盖子问题与证据不足的结论；补充检索；仍不足的标记为未验证。

7. **综合成文（Write）**
   按子问题组织结构；每条结论附来源编号；区分原文事实、跨源推论与模型判断。

8. **逐条引用核验**
   逐条打开引用，确认 URL 可达且原文片段支持对应论断；失败的删除或降级；统计引用可核查率。

9. **分级与审阅**
   结论分三档；由评审智能体或人工复核高影响结论与全部未验证结论。

10. **交付与归档**
    输出报告、证据表、引用库、覆盖范围说明与成本记录；归档研究计划以便后续更新。

## 质量标准（DoD）

引用与证据：

- [ ] 每条引用均来自检索，且 URL 可达
- [ ] 引用可核查率已统计并达到组织设定阈值
- [ ] 每条关键结论有至少两个独立来源，或已标注为单一来源
- [ ] 来源可信度已分级（一手 / 权威二手 / 一般二手 / 存疑）
- [ ] 引用元数据完整，缺失项显式标注「未知」

覆盖与平衡：

- [ ] 子问题清单全部覆盖，未覆盖项已说明原因
- [ ] 反方观点与冲突证据已独立呈现
- [ ] 未出现只呈现支持性证据的选择性偏倚
- [ ] 检索策略与查询词已记录，可复现

表达与分级：

- [ ] 事实、推论、判断已区分标注
- [ ] 结论已分三档：已验证 / 部分验证 / 未验证
- [ ] 未验证结论只列为假设，未混入结论
- [ ] 数值、日期、人名、机构名与来源一致

治理与成本：

- [ ] 未使用来源白名单之外的站点，或已获授权并留痕
- [ ] 成本在预算内，或已上报
- [ ] 研究计划、证据表与引用库已归档

## 常见失败与处理

| 失败现象 | 根因 | 处理方式 |
|---|---|---|
| 引用看起来规范但打不开 | 引用由模型生成 | 删除；重新检索；建立「引用必须来自检索」的硬约束 |
| 引用能打开但不支持该论断 | 只核对了 URL 未核对内容 | 逐条核对原文片段；不支持则删除或改写结论 |
| 报告很通顺但没有新信息 | 检索深度不足或只用头部结果 | 扩展查询词；检索一手来源；检查是否被少数来源主导 |
| 结论一边倒 | 未检索反方观点 | 强制设置反方检索轮次；独立呈现冲突 |
| 成本远超预期 | 未设预算或多智能体放大未计入 | 设置预算护栏；按子问题分批放行；记录 token 放大倍数 |
| 不同来源数字打架 | 口径或时点不同 | 并列呈现并注明口径与时点差异，不自行择一 |
| 检索很久找不到答案 | 问题过窄或术语不当 | 回到 Scope 步骤修正术语；或判定为无公开证据并如实说明 |
| 同一研究两次结论不同 | 检索随机性与来源漂移 | 固定来源快照与访问时间；记录检索式以便复现 |

## 示例

**任务**：梳理「科学数据汇交」相关的中国国家政策与标准，输出一份可内部传阅的合规要点清单。

1. 界定范围：时间范围 2018-01 至 2026-09；地域范围中国大陆；纳入国务院、部委与国家标准化管理机构发布的文件；排除地方规范性文件与研究论文（论文仅作解读参考）。
2. 计划卡点：范围与预算经合规负责人确认。
3. 并行检索：子问题拆为「法律与行政法规」「国务院规范性文件」「国家标准」「部门规章与机构办法」四条线；来源白名单限定为政府门户与标准管理机构站点。
4. 阅读摘录：抓取政策原文页面，摘录关键条款原文片段，记录 URL 与访问时间。
5. 交叉验证：每项条款寻找官方发布页与至少一份官方解读；记录表述差异。
6. 缺口识别：某三项国家标准未检索到标准全文页面，仅获得期刊论文的框架转述 → 标记为「条款原文未获取」。
7. 综合成文：按「法律—行政法规—规范性文件—国家标准—机构办法」组织，每条附来源编号。
8. 引用核验：逐条打开 URL；发现一处转述性来源无法访问 → 删除该引用，相关结论降级为部分验证。
9. 分级审阅：「先汇交、后验收」机制有官方原文支持，列为已验证；某标准的 7 阶段流程仅有论文转述，列为部分验证并注明「条款原文未获取」。
10. 交付归档：输出合规要点清单 + 证据表 + 引用库 + 覆盖范围说明（含未获取条款原文的声明）+ 成本记录。

4.3 Landing Checklist

4.3.1 Context Layer (L1)

• [ ] The research scope and sub-question list can be explicitly expressed and updated as progress is made
• [ ] Context isolation is adopted: sub-questions are assigned to independent sub-agents, compressed only at the aggregation layer
• [ ] The evidence table enters the context after sorting by relevance and deduplication
• [ ] Conflicting evidence and opposing viewpoints are not trimmed and are forcibly retained
• [ ] The glossary and domain background are retrievable

4.3.2 Tool & Execution Layer (L2)

• [ ] Search and browse tools support a source allowlist, enforced at the tool layer
• [ ] Every retrieval records the query, hit count, and adopted count
• [ ] The citation management structure supports URL, publication time, access time, and original excerpt
• [ ] Only registered MCP connectors are enabled
• [ ] Long pages support segmented reading

4.3.3 Orchestration & Control Layer (L3)

• [ ] The scope / research / write three-stage pipeline is implemented
• [ ] The orchestrator-worker architecture is available, and sub-agents can run in parallel
• [ ] The plan checkpoint takes effect before large-scale retrieval
• [ ] A supplementary retrieval loop after gap identification is supported
• [ ] Batched release is supported to control cost

4.3.4 Memory & State Layer (L4)

• [ ] The research plan can be persisted and resumed after interruption
• [ ] The citation library can be reused across tasks
• [ ] Failed search queries are recorded to avoid repeating ineffective retrievals
• [ ] Previous conclusions can be retrieved and timestamped

4.3.5 Evaluation & Observation Layer (L5)

• [ ] Citation verifiability rate is automatically tallied, and verification failures are automatically flagged
• [ ] The number of independent sources per key conclusion can be counted
• [ ] The presentation ratio of conflicting evidence can be checked
• [ ] External benchmarks (DRACO, SealQA-like) are used periodically for horizontal comparison, with results taken only as reference
• [ ] High-impact conclusions have a human review channel
• [ ] It is explicitly acknowledged that this direction lacks ground truth, and "accuracy N%" is not claimed

4.3.6 Governance & Security Layer (L6)

• [ ] The source allowlist is configurable and cannot be bypassed by agents
• [ ] Copyright and citation norms are internalized (length limits, source attribution)
• [ ] Cost guardrails are set (token, request-count, and amount caps)
• [ ] Permissions for retrieving restricted internal materials are isolated
• [ ] A human review checkpoint exists before external publication

5. Summary

The deep research direction shows a sharp two-sidedness: it is the direction where capability has been most thoroughly commoditized, and also the direction where ground truth is weakest.

Capability has been most thoroughly commoditized — from 2024-12 to 2026-07, within two years deep research completed the full evolution from a product button to a programmable API, from closed source to an open-source reference implementation, and from open retrieval to a restricted source allowlist. Any organization can call these capabilities directly today without building its own retrieval models.

Ground truth is weakest — a review has no standard answer. There is no result set to compare against, no physical law to constrain it, and no experiment to reproduce. This makes this direction the one with the highest hallucination-amplification risk: a fabricated citation enters the final report with full formatting, and without a line-by-line verification mechanism, no step intercepts it.

Therefore, the Harness proposition for this direction can be condensed into three points:

1. Citations can only come from retrieval, never from memory. This is the one red line with no exceptions.
2. Turn "undecidable correctness" into a "decidable provenance problem". We cannot automatically judge whether a conclusion is right, but we can automatically judge whether each citation actually exists and supports the corresponding claim. Citation verifiability rate thereby becomes the most valuable substitute metric in this direction.
3. Conflicting evidence must be presented. A review that presents only supportive evidence is more dangerous than having no review at all — it has complete citation formatting but no scientific honesty.

Finally, it bears repeating: this direction cannot claim an "accuracy N%" — both because there is no ground truth and because the mainstream vendors have not published independent reproducible accuracy benchmarks for their research APIs. Any such figure should be attributed to its source and have its limitations explained.

Information Gap Statement

1. There is no official or industry-recognized standard original text for the deep research direction's AGENTS.md / SKILL.md. Sections 4.1 and 4.2 are both standard drafts proposed in this article.
2. The phrasing "lead by 90.2%" for Anthropic's multi-agent research system is semantically ambiguous (the original is "90.2% ahead of a single Claude Opus 4 agent", likely a relative improvement rather than an absolute difference), marked [To be verified], and it is advisable to check the original Anthropic engineering blog before citing.
3. HLE 26.6% and 46.4% are figures mentioned by multiple sources with fairly high confidence, but official releases should be verified; BrowseComp 68.9% / 59.2%, DeepResearch Bench 46.5 / 48.9 pts, DeepSearchQA 66.1%, GPQA Diamond 94.3%, BrowseComp with full tools 85.9% (+45%) come from same-source secondary reporting and are marked [To be verified].
4. Pricing data changes frequently; the prices cited in this article are from the mid-2026 capture point: o3-deep-research is $10 per million input / $40 per million output tokens; ChatGPT Pro is $200/month; Google AI Pro is $19.99/month; Linkup is $0.25–2.50 each time; Perplexity Pro is $20/month and Max $200/month. The time point must be noted when citing.
5. SealQA-0 61% and DRACO 53.1 contain vendor self-reported components, are marked [To be verified], and are not an independent basis for capability comparison.
6. Neither OpenAI nor Gemini has published an independent reproducible accuracy benchmark for its research API — this is an objective fact of this direction's evaluation predicament, not an inference of this article.
7. The volume and page numbers of the original FAIR principles paper were not directly verified against the original page in this retrieval; this article only cites the FAIR principles themselves and the survey data and quotes in Springer Nature's official blog.
8. The runtime data (2–4 minutes, 5–30 minutes) comes from vendor and media descriptions and is a range description rather than precise measurement.
9. The policy and standards-sorting task in the Section 4.2 example is an illustrative construct, in which "the original text of certain three national standard clauses was not obtained" corresponds to the actual retrieval failure item in Section 11.3 of this report.

6. References

1. Deep Research Agents in Production 2026 (product chronicle and runtime characteristics) — Reactify Solutions. https://www.reactify-solutions.com/articles/deep-research-agents-production-2026
2. Best Deep Research API in 2026: OpenAI, Gemini and Linkup Compared (benchmark comparison and pricing) — Linkup. https://www.linkup.so/blog/best-deep-research-api-in-2026-openai-gemini-and-linkup-compared
3. AI Search Deep Research Comparison (runtime and pricing) — Fello AI. https://felloai.com/ai-search-deep-research-comparison/
4. Gemini Deep Research Agent vs OpenAI Deep Research (HLE, BrowseComp, DeepResearch Bench data, to be verified) — Agent Market Cap. https://agentmarketcap.ai/blog/2026/04/05/gemini-deep-research-agent-vs-openai-deep-research-autonomous-research-product-category
5. Gemini Deep Research Agent Preview (GPQA Diamond, BrowseComp with full tools data, to be verified) — Agent Market Cap. https://agentmarketcap.ai/blog/2026/04/05/gemini-deep-research-agent-preview-google-vs-openai-autonomous-research
6. Open Data: Progress, Challenges and Institutions (FAIR ten-year survey data and respondent quotes) — Springer Nature. https://springernature.com/gp/librarians/the-link/open-science-blogpost/open-data-progress-challenges-institutions/52378334
7. ACM Artifact Evaluation terminology definitions — arXiv 1908.05986. https://arxiv.org/pdf/1908.05986.pdf
8. Scientific Data Management Measures (Guo Ban Fa [2018] No. 17) — General Office of the State Council. https://www.gov.cn/zhengce/content/2018-04/02/content_5279272.htm
9. MCP scientific research laboratory tool ecosystem (literature and writing MCP Servers) — technical guide. https://chatforest.com/guides/mcp-scientific-research-laboratory/
10. MCP for Data Scientists: Notebooks (Jupyter / Colab MCP) — MCPFind. https://mcpfind.org/blog/mcp-for-data-scientists-notebooks
11. Agents for R&D Science (list of autonomous scientific discovery systems) — Orchestra Bio. https://orchestra.bio/blog/agents-for-r-d-science
12. Chinese Academy of Sciences scientific data management specification chain explanation — National Basic Disciplines Public Science Data Center. https://cms.casdc.cn/article/312