Back to Details

research

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Multi-Agent Research

Multi-Agent Research

Wave-based knowledge gathering with deferred synthesis. Mined from 300+ real research dispatches — the pattern that consistently produces actionable intelligence.
Core insight: Research breadth-first, synthesize after. Don't draw conclusions from the first 3 results. Deploy agents in waves, accumulate findings, then synthesize with the full picture.
基于波次调度、延迟合成的知识收集方法，提炼自300多份真实调研任务输出——该模式可稳定输出可落地的情报结论。
核心思路： 先开展广度优先调研，再进行结果合成。不要仅根据前3条结果就下结论。分波次部署Agent，积累调研发现，再基于完整信息合成结论。

The Process

流程

dot
digraph research {
    rankdir=TB;
    node [shape=box];

    "1. PRIME" [style=filled, fillcolor="#e8e8ff"];
    "2. WAVE 1: Broad Sweep" [style=filled, fillcolor="#ffe8e8"];
    "3. GAP ANALYSIS" [style=filled, fillcolor="#fff8e0"];
    "4. WAVE 2+: Targeted" [style=filled, fillcolor="#ffe8e8"];
    "5. SYNTHESIZE" [style=filled, fillcolor="#e8ffe8"];
    "6. DECIDE & RECORD" [style=filled, fillcolor="#e8e8ff"];

    "1. PRIME" -> "2. WAVE 1: Broad Sweep";
    "2. WAVE 1: Broad Sweep" -> "3. GAP ANALYSIS";
    "3. GAP ANALYSIS" -> "4. WAVE 2+: Targeted";
    "4. WAVE 2+: Targeted" -> "3. GAP ANALYSIS" [label="still gaps", style=dashed];
    "3. GAP ANALYSIS" -> "5. SYNTHESIZE" [label="coverage sufficient"];
    "5. SYNTHESIZE" -> "6. DECIDE & RECORD";
}
dot
digraph research {
    rankdir=TB;
    node [shape=box];

    "1. PRIME" [style=filled, fillcolor="#e8e8ff"];
    "2. WAVE 1: Broad Sweep" [style=filled, fillcolor="#ffe8e8"];
    "3. GAP ANALYSIS" [style=filled, fillcolor="#fff8e0"];
    "4. WAVE 2+: Targeted" [style=filled, fillcolor="#ffe8e8"];
    "5. SYNTHESIZE" [style=filled, fillcolor="#e8ffe8"];
    "6. DECIDE & RECORD" [style=filled, fillcolor="#e8e8ff"];

    "1. PRIME" -> "2. WAVE 1: Broad Sweep";
    "2. WAVE 1: Broad Sweep" -> "3. GAP ANALYSIS";
    "3. GAP ANALYSIS" -> "4. WAVE 2+: Targeted";
    "4. WAVE 2+: Targeted" -> "3. GAP ANALYSIS" [label="still gaps", style=dashed];
    "3. GAP ANALYSIS" -> "5. SYNTHESIZE" [label="coverage sufficient"];
    "5. SYNTHESIZE" -> "6. DECIDE & RECORD";
}

Phase 1: PRIME

阶段1：PRIME（准备阶段）

Search what you already know before spawning a single agent.
在生成任何Agent之前，先检索你已有的知识库。

Actions

操作步骤

  1. Search Sibyl first:
    sibyl search "<research topic>"
sibyl search "<related technology>"
sibyl search "<prior decision in this area>"
  2. Check for stale knowledge:
    • If Sibyl has entries older than 3 months on fast-moving topics (frameworks, models, cloud services) → research anyway, but use existing knowledge as baseline
    • If Sibyl has recent entries → present them first, ask if deeper research is needed
  3. Define the research question clearly:
    • NOT: "research databases" (too vague)
    • YES: "compare PostgreSQL vs CockroachDB for multi-region write-heavy workloads with <10ms p99 latency requirement"
  4. Set the research budget:
    DepthAgentsTimeWhen
    Quick scan2-32-5 minKnown domain, just need latest info
    Standard5-1010-15 minTechnology evaluation, architecture options
    Deep dive10-3020-40 minGreenfield decisions, SOTA analysis
    Exhaustive30-60+40-90 minNew project inception, competitive landscape
  1. 优先检索Sibyl：
    sibyl search "<research topic>"
sibyl search "<related technology>"
sibyl search "<prior decision in this area>"
  2. 检查知识是否过期：
    • 如果Sibyl中快速迭代领域（框架、模型、云服务）的条目超过3个月 → 仍需开展调研，但可将现有知识作为基线
    • 如果Sibyl中有近期条目 → 优先展示现有内容，询问是否需要更深度的调研
  3. 清晰定义调研问题：
    • 反面示例："调研数据库"（过于模糊）
    • 正面示例："对比PostgreSQL和CockroachDB在多区域写密集、p99延迟要求<10ms的工作负载下的表现"
  4. 设定调研预算：
    调研深度调用Agent数量耗时适用场景
    快速扫描2-32-5分钟已知领域，仅需获取最新信息
    标准调研5-1010-15分钟技术评估、架构选型
    深度调研10-3020-40分钟新领域决策、SOTA分析
    穷尽调研30-60+40-90分钟新项目启动、竞争格局分析

Phase 2: WAVE 1 — Broad Sweep

阶段2：第一波次——广度覆盖

Deploy the first wave of agents across the full research surface.
在全调研范围内部署第一波Agent。

Agent Design Principles

Agent设计原则

Each agent gets:
  • One specific topic (not "research everything about X")
  • An output file path (no ambiguity about where to write)
  • Search hints (include year: "search [topic] 2026")
  • 8-12 numbered coverage items (scope the research precisely)
  • Source quality guidance ("prefer official docs and GitHub repos over blog posts")
每个Agent分配：
  • 一个明确的细分主题（不要是"调研X的所有内容"）
  • 一个输出文件路径（避免输出位置歧义）
  • 搜索提示（包含年份："search [topic] 2026"）
  • 8-12个编号覆盖项（精准限定调研范围）
  • 来源质量指引（"优先参考官方文档和GitHub仓库，其次是博客文章"）

Wave 1 Template

第一波次Prompt模板

markdown
Research [SPECIFIC_TOPIC] for [PROJECT/DECISION].

Create a research doc at docs/research/[filename].md covering:

1. Current state (latest version, recent changes)
2. [Specific capability A relevant to our use case]
3. [Specific capability B]
4. [Integration with our stack: list specific technologies]
5. Performance characteristics / benchmarks
6. Known limitations and gotchas
7. Community health (stars, activity, maintenance)
8. Comparison with alternatives (name 2-3 specific alternatives)

Use WebSearch for current information. Include dates on all facts.
Cite sources with URLs.
markdown
Research [SPECIFIC_TOPIC] for [PROJECT/DECISION].

Create a research doc at docs/research/[filename].md covering:

1. Current state (latest version, recent changes)
2. [Specific capability A relevant to our use case]
3. [Specific capability B]
4. [Integration with our stack: list specific technologies]
5. Performance characteristics / benchmarks
6. Known limitations and gotchas
7. Community health (stars, activity, maintenance)
8. Comparison with alternatives (name 2-3 specific alternatives)

Use WebSearch for current information. Include dates on all facts.
Cite sources with URLs.

Deployment Rules

部署规则

  • ALL Wave 1 agents run in background — no dependencies between them
  • 3-4 seconds between dispatches — avoid rate limiting
  • Each agent writes its own file — no shared outputs
  • Group by theme: If researching 12 topics, group into 3-4 thematic clusters
  • 所有第一波Agent都在后台运行 — 互相之间无依赖
  • 每次调度间隔3-4秒 — 避免触发限流
  • 每个Agent写入独立文件 — 无共享输出
  • 按主题分组： 如果要调研12个主题，可划分为3-4个主题组

Coverage Strategy

覆盖策略

For technology evaluations, cover these dimensions:
DimensionQuestion
CapabilityDoes it do what we need?
PerformanceIs it fast enough?
EcosystemDoes it integrate with our stack?
MaturityIs it production-ready?
CommunityWill it be maintained in 2 years?
CostWhat does it cost at our scale?
MigrationHow hard is it to adopt/abandon?
针对技术评估场景，需覆盖以下维度：
维度核心问题
功能能力是否能满足我们的需求？
性能表现速度是否达标？
生态适配是否能和我们的技术栈集成？
成熟度是否可用于生产环境？
社区活跃度未来2年是否会持续维护？
成本在我们的规模下使用成本是多少？
迁移成本接入/放弃该技术的难度有多高？

Phase 3: GAP ANALYSIS

阶段3：缺口分析

After Wave 1 completes, identify what's missing before synthesizing.
第一波次完成后，在合成结论前先识别信息缺口。

Actions

操作步骤

  1. Read all Wave 1 outputs — skim each research doc
  2. Identify gaps:
    • Dimensions not covered?
    • Contradictory findings between agents?
    • Questions raised but not answered?
    • Missing comparisons?
  3. Check for bias:
    • Are all findings positive? (Suspicious — look for failure cases)
    • Did agents only find official docs? (Need community/real-world experience)
    • Same sources cited by multiple agents? (Need diversity)
  1. 通读所有第一波次输出 — 快速浏览每份调研文档
  2. 识别信息缺口：
    • 是否有维度未覆盖？
    • 不同Agent的调研结果是否存在矛盾？
    • 是否有提出但未解答的问题？
    • 是否缺少对比内容？
  3. 检查偏差：
    • 所有结论都是正面的？（可信度低 — 需查找失败案例）
    • Agent是否只找到了官方文档？（需要社区/实际使用经验的信息）
    • 多个Agent引用了相同的来源？（需要来源多样性）

Decision Point

决策节点

FindingAction
Good coverage, minor gapsSynthesize now, note gaps
Significant gapsDeploy Wave 2 targeted agents
Contradictory findingsDeploy verification agents to resolve
Entirely new direction emergedDeploy Wave 2 in new direction
发现下一步动作
覆盖度良好，仅存在少量缺口立即合成结论，标注存在的缺口
存在显著缺口部署第二波定向调研的Agent
调研结果存在矛盾部署验证Agent解决冲突
出现了全新的调研方向面向新方向部署第二波Agent

Phase 4: WAVE 2+ — Targeted Research

阶段4：第2+波次——定向调研

Fill specific gaps identified in the analysis.
补充缺口分析中识别的特定信息缺口。

Wave 2 Agents Are Different

第二波次Agent的差异

  • Smaller scope — one specific question per agent
  • Higher quality bar — "find production experience reports, not just docs"
  • Cross-reference — "Agent X found [claim]. Verify this against [alternative source]"
  • Deep reads — "Read the full README and API docs for [library], not just the landing page"
  • 范围更小 — 每个Agent仅负责一个特定问题
  • 质量要求更高 — "查找生产实践报告，不仅仅是官方文档"
  • 交叉验证 — "Agent X得出了[结论]，请从[其他来源]验证该结论"
  • 深度阅读 — "阅读[库]的完整README和API文档，不仅仅是首页介绍"

When to Stop

停止调研的条件

Stop deploying waves when:
  • The research question can be answered with confidence
  • Additional agents would produce diminishing returns
  • You have >= 2 independent sources for key claims
  • The user signals "enough, let's decide"
Max 3 waves for most research. If 3 waves haven't answered the question, the question needs reframing.
满足以下条件时停止部署新的波次：
  • 可以有信心地回答调研问题
  • 新增Agent带来的信息价值边际递减
  • 核心结论有≥2个独立来源支撑
  • 用户明确表示"足够了，开始决策"
绝大多数调研最多开展3波次。 如果3波次后仍无法解答问题，说明问题本身需要重新梳理。

Phase 5: SYNTHESIZE

阶段5：合成结论

Combine all findings into actionable intelligence. This is where the magic happens.
将所有调研发现整合为可落地的情报，这是最核心的环节。

Synthesis Structure

结论合成结构

markdown
undefined
markdown
undefined

Research: [Topic]

Research: [Topic]

TL;DR

TL;DR

[2-3 sentences. The answer, not the journey.]
[2-3句话，直接给出答案，而非描述调研过程。]

Recommendation

Recommendation

[Clear choice with justification. Don't hedge — pick one.]
[给出明确的选择和理由，不要模棱两可，确定一个最优选项。]

Options Evaluated

Options Evaluated

OptionFitMaturityPerfEcosystemVerdict
A............Best for [X]
B............Best for [Y]
C............Avoid: [reason]
OptionFitMaturityPerfEcosystemVerdict
A............Best for [X]
B............Best for [Y]
C............Avoid: [reason]

Key Findings

Key Findings

  1. [Most important finding with source]
  2. [Second most important]
  3. [Third most important]
  1. [最重要的发现，附带来源]
  2. [第二重要的发现]
  3. [第三重要的发现]

Risks & Gotchas

Risks & Gotchas

  • [Known issue or limitation]
  • [Migration complexity]
  • [Hidden cost]
  • [已知问题或限制]
  • [迁移复杂度]
  • [隐藏成本]

Sources

Sources

  • Source 1 — [what it contributed]
  • Source 2 — [what it contributed]
undefined
  • Source 1 — [贡献的信息]
  • Source 2 — [贡献的信息]
undefined

Synthesis Rules

合成规则

  1. Lead with the recommendation. Don't make the reader wade through findings to find the answer.
  2. Separate facts from opinions. "PostgreSQL supports JSONB" (fact) vs "PostgreSQL is better for this use case" (opinion backed by evidence).
  3. Include dissenting evidence. If one source contradicts the recommendation, say so. Don't cherry-pick.
  4. Date everything. "As of Feb 2026, [library] is at v4.2" — research spoils fast.
  5. Note confidence level. "High confidence: well-documented. / Low confidence: based on one blog post."
  1. 开头直接给出推荐结论。 不要让读者翻完所有内容才找到答案。
  2. 区分事实和观点。 例如"PostgreSQL支持JSONB"是事实，"PostgreSQL更适合该场景"是有证据支撑的观点。
  3. 包含相反的证据。 如果有来源和推荐结论矛盾，要明确说明，不要选择性忽略。
  4. 所有结论标注时间。 例如"截至2026年2月，[库]的最新版本为v4.2" — 调研内容的时效性很强。
  5. 标注置信度。 例如"高置信度：文档完善。/ 低置信度：仅基于单篇博客文章。"

Phase 6: DECIDE & RECORD

阶段6：决策与记录

Lock in the decision and capture it for future sessions.
确定最终决策并记录下来，供后续使用。

Actions

操作步骤

  1. Present the synthesis to the user with a clear recommendation
  2. Record in Sibyl:
    sibyl add "Research: [topic]" "Evaluated [options]. Chose [X] because [reasons]. Key risk: [Y]. Sources: [primary URLs]. Date: [today]."
  3. Archive research docs — keep the wave outputs for reference:
    • If in a project: 
      docs/research/[topic]/
    • If general knowledge: Sibyl learning entry is sufficient
  4. Exit to next action:
    Next StepWhen
    /hyperskills:brainstorm
    		Research surfaced multiple viable approaches
    /hyperskills:plan
    		Decision made, ready to decompose implementation
    /hyperskills:orchestrate
    		Decision made, work is parallelizable
    Direct implementationResearch confirmed a simple path
  1. 向用户展示合成后的结论，附带明确的推荐意见
  2. 记录到Sibyl中：
    sibyl add "Research: [topic]" "Evaluated [options]. Chose [X] because [reasons]. Key risk: [Y]. Sources: [primary URLs]. Date: [today]."
  3. 归档调研文档 — 保留各波次的输出供后续参考：
    • 项目相关的调研：存入
      docs/research/[topic]/
      目录
    • 通用知识：仅在Sibyl中记录学习条目即可
  4. 进入下一步动作：
    下一步指令适用场景
    /hyperskills:brainstorm
    		调研发现了多个可行的实现路径
    /hyperskills:plan
    		已完成决策，准备拆解执行步骤
    /hyperskills:orchestrate
    		已完成决策，工作可并行推进
    直接开始实现调研确认路径非常简单

Quick Research Mode

快速调研模式

For focused questions that don't need the full wave protocol:
  1. Search Sibyl (always)
  2. 2-3 targeted searches (WebSearch + WebFetch on key URLs)
  3. Synthesize inline (no separate docs)
  4. Record if non-obvious (Sibyl learning)
Use when: "What's the latest version of X?", "Does Y support Z?", "What's the recommended way to do W?"
针对不需要完整波次流程的聚焦型问题：
  1. 检索Sibyl（必做步骤）
  2. 2-3次定向搜索（WebSearch + 关键URL的WebFetch）
  3. 直接合成结果（不需要单独生成文档）
  4. 非直观结论进行记录（存入Sibyl学习库）
适用场景： "X的最新版本是什么？"、"Y是否支持Z？"、"实现W的推荐方案是什么？"

Research Patterns by Type

不同类型调研的落地模式

Technology Evaluation

技术评估

Wave 1: Official docs + GitHub README for each option (parallel)
Wave 2: Production experience + benchmarks (parallel)
Synthesize: Comparison matrix + recommendation
Wave 1: Official docs + GitHub README for each option (parallel)
Wave 2: Production experience + benchmarks (parallel)
Synthesize: Comparison matrix + recommendation

Codebase Archaeology

代码库考古

Wave 1: Explore agents mapping each subsystem (parallel)
Wave 2: Grep for specific patterns / usage (parallel)
Synthesize: Architecture diagram + dependency map
Wave 1: Explore agents mapping each subsystem (parallel)
Wave 2: Grep for specific patterns / usage (parallel)
Synthesize: Architecture diagram + dependency map

SOTA Analysis

SOTA分析

Wave 1: WebSearch for latest papers, blog posts, releases (parallel)
Wave 2: Deep read the most relevant 3-5 sources (parallel)
Synthesize: What's genuinely novel vs rehashed + recommendation
Wave 1: WebSearch for latest papers, blog posts, releases (parallel)
Wave 2: Deep read the most relevant 3-5 sources (parallel)
Synthesize: What's genuinely novel vs rehashed + recommendation

Competitive Landscape

竞争格局分析

Wave 1: Feature matrix for each competitor (parallel)
Wave 2: Pricing, community size, trajectory (parallel)
Synthesize: Positioning matrix + gap analysis
Wave 1: Feature matrix for each competitor (parallel)
Wave 2: Pricing, community size, trajectory (parallel)
Synthesize: Positioning matrix + gap analysis

Anti-Patterns

反模式

Anti-PatternFix
Synthesizing after Wave 1 onlyWait for gap analysis — premature conclusions miss nuance
50 agents with "research everything"Specific scope per agent — vague prompts produce vague results
Only official documentationInclude community experience — docs show intent, community shows reality
No dates on findingsDate everything — research spoils faster than produce
No recommendationForce a decision — "more research needed" is only valid with a specific question
Researching what Sibyl already knowsAlways prime first — don't burn tokens re-discovering known patterns
反模式解决方案
仅完成第一波次就合成结论先完成缺口分析再合成 — 过早下结论会遗漏细节
50个Agent都分配"调研所有内容"的任务每个Agent配置明确的范围 — 模糊的Prompt只会产出模糊的结果
仅参考官方文档纳入社区使用经验 — 文档描述的是预期，社区反馈的是实际情况
结论没有标注时间所有结论都标注时间 — 调研内容的过期速度比生鲜食品还快
没有给出推荐结论强制给出决策 — "需要更多调研"仅在有明确后续问题时才有效
重复调研Sibyl已有的知识永远先做准备阶段 — 不要浪费算力重新发现已知的内容

What This Skill is NOT

该技能的边界

  • Not a substitute for reading code. If the answer is in the codebase, read the codebase.
  • Not an infinite loop. Max 3 waves. If that's not enough, reframe the question.
  • Not required for known domains. If you already know the answer, just say so and cite your knowledge.
  • Not a delay tactic. Research serves a decision. If no decision follows, the research was waste.
  • 不能替代代码阅读。 如果答案就在代码库中，直接阅读代码即可。
  • 不是无限循环。 最多开展3波次调研，如果不够就重新梳理问题。
  • 已知领域不需要使用。 如果你已经知道答案，直接给出结论并标注知识来源即可。
  • 不是拖延手段。 调研是为了支撑决策，如果没有后续决策，调研就是浪费资源。