brycewang-stanford/skills/05-kthorn-research-superpower/research/subagent-driven-review

<!-- ╔══════════════════════════════════════════════════════════════╗ ║ 本文件为开源 Skill 原始文档，收录仅供学习与研究参考 ║ ║ CoPaper.AI 收集整理 | https://copaper.ai ║ ╚══════════════════════════════════════════════════════════════╝ 来源仓库: https://github.com/kthorn/research-superpower 项目名称: research-superpower 开源协议: MIT License 收录日期: 2026-04-02 声明: 本文件版权归原作者所有。此处收录旨在为社会科学实证研究者 提供 AI Agent Skills 的集中参考。如有侵权，请联系删除。 -->

---

name: Subagent-Driven Literature Review description: Use parallel subagents for large-scale paper screening and deep dive analysis when_to_use: Large literature searches (50+ papers), parallel paper screening, deep dive analysis on multiple papers, citation network exploration, when main context is getting full version: 1.0.0

Subagent-Driven Literature Review

Overview

Core principle: Fresh subagent per batch + consolidation between batches = fast parallel screening with quality control

For large literature reviews (50+ papers), dispatching parallel or sequential subagents dramatically speeds up screening while maintaining quality through consolidation checkpoints.

When to Use

Use subagent-driven approach when:

• Large searches: 50+ papers to screen
• Parallelizable work: Papers are independent, can be screened separately
• Deep dive tasks: Multiple papers need detailed extraction (data tables, methods, datasets)
• Citation exploration: Following citation networks recursively
• Context management: Main context getting full, need fresh context
• Time pressure: Need results faster than sequential screening

Do NOT use when:

• Small searches (<20 papers) - overhead not worth it
• Need real-time user visibility into every paper
• Papers require cross-comparison during screening
• Simple, fast screening tasks

Use Cases

1. Parallel Paper Screening (Most Common)

Scenario: You have 100 papers from PubMed search to screen for relevance

Pattern:

Main agent:
1. Splits 100 papers into 5 batches of 20
2. Dispatches 5 subagents IN PARALLEL (single message, multiple Task calls)
3. Each subagent:
   - Fetches abstracts for its batch
   - Scores using rubric
   - Returns JSON with results
4. Main agent consolidates results into papers-reviewed.json

Time savings: 5x faster than sequential!

Prompt template for subagent:

I need you to screen papers 1-20 from this PMID list for relevance to [QUERY].

PMIDs to screen: [PMID list]

Use the evaluating-paper-relevance skill to:
1. Fetch abstract for each PMID
2. Score 0-10 based on:
   - Keywords: [list]
   - Data types needed: [measurements, protocols, datasets, etc.]
3. Return JSON:

{
  "screened_papers": [
    {"pmid": "12345", "score": 8, "status": "relevant", "reason": "..."},
    ...
  ],
  "stats": {"highly_relevant": 3, "relevant": 5, "not_relevant": 12}
}

Do NOT update papers-reviewed.json - return results only.

**Rate limiting (CRITICAL - PubMed limits are SHARED across all parallel subagents):**
- If you are the ONLY subagent running: Use 500ms delays (2 req/sec, safe)
- If running with OTHER parallel subagents: Use longer delays to share capacity
  - You are 1 of 2 parallel: Use 1 second delays
  - You are 1 of 3 parallel: Use 1.5 second delays
  - You are 1 of 5 parallel: Use 2.5 second delays
- If you get HTTP 429 errors: Wait 5 seconds, then use 5-second delays for remaining requests

2. Deep Dive on Priority Papers

Scenario: Initial screening identified 15 highly relevant papers, need detailed data extraction from each

Pattern:

Main agent:
1. Creates TodoWrite with 15 tasks (one per paper)
2. For each paper, dispatches subagent to:
   - Fetch full text (PMC, Unpaywall)
   - Extract relevant data (tables, figures, methods)
   - Identify key findings
   - Return structured findings
3. Main agent consolidates into SUMMARY.md
4. Reviews and adds to papers-reviewed.json

Can dispatch in parallel (5 at a time) or sequentially

Prompt template for subagent:

Deep dive analysis for paper PMID [12345] / DOI [10.xxxx/yyyy]

Use evaluating-paper-relevance skill to:
1. Check for curated data sources (if applicable to domain)
2. Fetch full text (try PMC, then Unpaywall if paywalled)
3. Extract relevant data based on research domain:
   - Data tables and measurements
   - Methods and protocols
   - Key results and findings
   - Figures with relevant information
4. Return structured JSON:

{
  "pmid": "12345",
  "doi": "10.xxxx/yyyy",
  "full_text_source": "PMC" or "Unpaywall" or "paywalled",
  "data_sources": ["Table 1", "Figure 3", "Supplementary Data"],
  "key_measurements": ["specific values or ranges found"],
  "methods_summary": "Brief description of methods",
  "key_findings": ["Finding 1", "Finding 2", ...],
  "data_availability": "GEO: GSE12345" or "Code: github.com/..." or null
}

Do NOT update papers-reviewed.json - return findings only.

3. Citation Network Exploration

Scenario: Found one highly relevant paper, need to explore forward and backward citations

Pattern:

Main agent:
1. Dispatches two subagents IN PARALLEL:
   - Subagent A: Fetch and screen forward citations
   - Subagent B: Fetch and screen backward citations
2. Each returns list of promising PMIDs with scores
3. Main agent:
   - Consolidates results
   - Removes duplicates
   - Adds to screening queue
   - Updates papers-reviewed.json

Prompt template for subagent:

Find and screen forward citations for PMID [12345].

Use traversing-citations skill to:
1. Fetch forward citations from PubMed or OpenCitations
2. Screen abstracts for relevance to [QUERY]
3. Score each citation (0-10)
4. Return JSON with promising papers (score ≥7):

{
  "seed_pmid": "12345",
  "direction": "forward",
  "citations_found": 45,
  "relevant_citations": [
    {"pmid": "67890", "score": 8, "title": "...", "reason": "..."},
    ...
  ]
}

Do NOT update papers-reviewed.json - return results only.

4. Domain-Specific Extraction

Examples by domain:

Genomics:

Subagent extracts:
- GEO/SRA/ENA accessions
- Sample sizes and conditions
- Sequencing methods (RNA-seq, WGS, etc.)
- Analysis pipelines
- Differential expression results

Computational methods:

Subagent extracts:
- Algorithm descriptions
- Code repositories (GitHub, GitLab, etc.)
- Benchmark datasets used
- Performance metrics
- Implementation details

Clinical research:

Subagent extracts:
- Study design (RCT, cohort, etc.)
- Sample size and demographics
- Intervention details
- Primary outcomes
- Statistical methods

Ecology/Environmental:

Subagent extracts:
- Study sites and coordinates
- Sampling methods
- Species/taxa studied
- Environmental measurements
- Data repositories

Workflow: Parallel Screening

Step 1: Plan and Split

Main agent tasks:

1. Load PMID list from search results
2. Decide on batch size (typically 15-25 papers per subagent)
3. Create TodoWrite with batches
4. Prepare subagent prompts

Example TodoWrite:

- Screen papers batch 1 (PMIDs 1-20)
- Screen papers batch 2 (PMIDs 21-40)
- Screen papers batch 3 (PMIDs 41-60)
- Screen papers batch 4 (PMIDs 61-80)
- Screen papers batch 5 (PMIDs 81-100)
- Consolidate all subagent results
- Generate SUMMARY.md from consolidated data

Step 2: Dispatch Subagents

CRITICAL: Dispatch all subagents in PARALLEL using single message with multiple Task calls

Example:

I'm dispatching 5 subagents in parallel to screen 100 papers.

[Uses Task tool 5 times in single message]

Why parallel: 5x speed improvement vs sequential!

Step 3: Collect Results

Main agent:

1. Wait for all subagents to complete
2. Collect JSON results from each
3. Validate format and completeness

Check for:

• All PMIDs were screened
• Scoring rubric was applied consistently
• No papers missing

Step 4: Consolidate

Main agent:

1. Merge all subagent results
2. Remove duplicates (if any overlap between batches)
3. Sort by relevance score
4. Add ALL papers to papers-reviewed.json:

{
  "10.1234/example.2023": {
    "pmid": "12345",
    "status": "highly_relevant",
    "score": 9,
    "source": "pubmed_search_batch1",
    "screened_by": "subagent",
    "timestamp": "2025-10-11T14:30:00Z",
    "found_data": ["measurements", "methods", "datasets"]
  }
}

Mark source as "subagent" or "pubmed_search_batch1" etc.

Step 5: Review Quality

Main agent checks:

• Scoring appears consistent across batches
• No batch has dramatically different hit rate (could indicate problem)
• Highly relevant papers make sense
• Any papers needing manual re-review?

Red flags:

• One batch found 10 relevant papers, others found 0-1 (inconsistent scoring?)
• Papers marked "highly relevant" don't match keywords
• Missing expected papers

If issues found: Re-screen problematic batch manually or with fresh subagent

Step 6: Generate Summary

Main agent:

1. Create SUMMARY.md with all highly relevant and relevant papers
2. Sort by score
3. Add statistics
4. Note which papers need deep dive

Step 7: Optional Deep Dive

For highly relevant papers (score ≥8):

Option A: Dispatch subagents sequentially

For each highly relevant paper:
  - Dispatch one subagent per paper
  - Subagent does deep dive extraction
  - Main agent consolidates findings immediately
  - Updates SUMMARY.md progressively

Option B: Dispatch subagents in parallel batches

Batch 1: Papers 1-5 (dispatch 5 subagents in parallel)
Wait for completion, consolidate
Batch 2: Papers 6-10 (dispatch 5 subagents in parallel)
Wait for completion, consolidate
...

Workflow: Citation Exploration

Step 1: Identify Seed Papers

Find 2-3 highly relevant papers from initial screening

Step 2: Dispatch Citation Subagents

For each seed paper, dispatch TWO subagents in parallel:

1. Forward citations (who cited this paper?)
2. Backward citations (what did this paper cite?)

Prompt each subagent with:

• Seed PMID
• Relevance criteria
• Return only papers scoring ≥7

Step 3: Consolidate Citations

Main agent:

1. Collects all citation results
2. Removes duplicates
3. Removes papers already in papers-reviewed.json
4. Creates new screening queue

Step 4: Screen New Papers

Option A: Dispatch new batch screening subagents for citation results Option B: Main agent screens smaller batch manually

Step 5: Iterate

If citation exploration found many new relevant papers:

• Consider exploring citations from those papers too
• Be careful of exponential growth!
• Set stopping criteria (e.g., max 3 levels deep, max 200 total papers)

Integration with Other Skills

Works with:

• evaluating-paper-relevance: Subagents use this for individual paper screening
• traversing-citations: Subagents use this for citation exploration
• finding-open-access-papers: Subagents check Unpaywall for paywalled papers
• checking-chembl: Subagents can check curated databases (when applicable)

Combines with:

• writing-plans: Create screening plan before dispatching subagents
• TodoWrite: Track batches and consolidation progress

Consolidation Patterns

Pattern 1: JSON Aggregation

Subagents return structured JSON, main agent merges:

# Pseudo-code for consolidation
all_results = []
for subagent_output in subagent_results:
    results = parse_json(subagent_output)
    all_results.extend(results['screened_papers'])

# Sort by score
all_results.sort(key=lambda x: x['score'], reverse=True)

# Update papers-reviewed.json
for paper in all_results:
    papers_reviewed[paper['doi']] = {
        'pmid': paper['pmid'],
        'status': paper['status'],
        'score': paper['score'],
        'source': f"subagent_batch_{paper['batch_id']}",
        'timestamp': now()
    }

Pattern 2: Progressive Consolidation

Consolidate after each subagent completes (sequential dispatch):

Dispatch subagent 1 → wait → consolidate → dispatch subagent 2 → wait → consolidate → ...

Advantage: See progress incrementally Disadvantage: Slower than full parallel

Pattern 3: Batch Consolidation

Dispatch N subagents in parallel, consolidate batch, repeat:

Dispatch 5 subagents → wait for all 5 → consolidate → dispatch next 5 → ...

Advantage: Balance between speed and manageable consolidation Disadvantage: More complex than full parallel or sequential

Common Mistakes

Not dispatching in parallel: Sending Task calls sequentially wastes time → Use single message with multiple Task calls Subagents updating tracking files: Causes conflicts → Subagents return JSON only, main agent updates files Inconsistent scoring: Different subagents use different rubrics → Provide clear rubric in prompt No quality review: Blindly trusting subagent results → Always review consolidated results Too many parallel subagents: Dispatching 20+ at once → Keep to 5-10 parallel max Forgetting rate limits: Subagents hit API limits → Include rate limiting in prompts (500ms for single agent, 2.5 seconds for 5 parallel agents) No source tracking: Can't tell which batch found which papers → Add batch_id or source field Duplicate work: Multiple subagents screen same papers → Carefully split PMID lists with no overlap

Cost Considerations

Subagent usage has cost implications:

Token usage per subagent:

• Screening 20 papers: ~10-15K tokens per subagent
• Deep dive 1 paper: ~5-10K tokens per subagent
• Citation exploration: ~8-12K tokens per subagent

Trade-off:

• Parallel screening: Higher cost, much faster (5x speed)
• Sequential screening: Lower cost, slower
• Consider for time-sensitive research

Cost-saving strategies:

• Use subagents for large batches only (50+ papers)
• Screen manually for small searches (<20 papers)
• Parallel dispatch for initial screening (speed matters)
• Sequential dispatch for deep dive (can review progressively)

Examples

Example 1: Screen 80 Papers in Parallel

Initial state: Have 80 PMIDs from PubMed search

Main agent:

I'll dispatch 4 subagents in parallel to screen these 80 papers.

Batch 1: PMIDs 1-20 [dispatches subagent with prompt]
Batch 2: PMIDs 21-40 [dispatches subagent with prompt]
Batch 3: PMIDs 41-60 [dispatches subagent with prompt]
Batch 4: PMIDs 61-80 [dispatches subagent with prompt]

[Uses Task tool 4 times in single message]

Now waiting for all subagents to complete...

After subagents complete:

All 4 subagents have completed. Consolidating results:

Batch 1: 3 highly relevant, 5 relevant, 12 not relevant
Batch 2: 2 highly relevant, 7 relevant, 11 not relevant
Batch 3: 4 highly relevant, 6 relevant, 10 not relevant
Batch 4: 1 highly relevant, 4 relevant, 15 not relevant

Total: 10 highly relevant, 22 relevant, 48 not relevant

Updating papers-reviewed.json with all 80 papers...
Generating SUMMARY.md with 32 relevant papers...

Next: Deep dive on 10 highly relevant papers?

Example 2: Deep Dive on 12 Papers

Initial state: Have 12 highly relevant papers needing data extraction

Main agent:

I'll dispatch 12 subagents (in 3 batches of 4) to do deep dive analysis.

Batch 1 papers: PMID 12345, 23456, 34567, 45678
[Dispatches 4 subagents in parallel with deep dive prompts]

Waiting for batch 1 to complete...
[Consolidates batch 1 results into SUMMARY.md]

Batch 2 papers: PMID 56789, 67890, 78901, 89012
[Dispatches 4 subagents in parallel]
...

Result: All 12 papers analyzed with structured data in ~10-15 minutes vs 1-2 hours sequential

Example 3: Citation Exploration

Initial state: Found key paper PMID 12345

Main agent:

I'll explore citations from this key paper.

Dispatching 2 subagents in parallel:
1. Forward citations (who cited PMID 12345?)
2. Backward citations (what did PMID 12345 cite?)

[Uses Task tool twice in single message]

Waiting for citation exploration...

Forward citations: Found 34 citations, 8 appear relevant
Backward citations: Found 42 references, 6 appear relevant

New papers to screen: 14 (after removing duplicates)

Now dispatching screening subagent for these 14 papers...

Quick Reference

| Task | Subagent Pattern | Parallel? | Consolidation | |------|-----------------|-----------|---------------| | Screen 100 papers | 5 batches of 20 | Yes (5 parallel) | Merge JSON, update papers-reviewed.json | | Deep dive on 15 papers | 15 individual tasks | Yes (batches of 5) | Add findings to SUMMARY.md progressively | | Citation exploration | 2-3 citation tasks | Yes | Merge, dedupe, add to screening queue | | Data extraction | 1 per paper | Sequential or batched | Update papers-reviewed.json with findings |

Decision Tree

Have literature review task?
├─ <20 papers?
│  └─ Screen manually (no subagents)
├─ 20-50 papers?
│  ├─ Time-sensitive? → Use subagents (2-3 batches)
│  └─ Not urgent? → Screen manually
└─ 50+ papers?
   ├─ Initial screening → Use parallel subagents (5-10 batches)
   ├─ Deep dive needed? → Use sequential or batched subagents
   └─ Citation exploration? → Use parallel subagents per seed paper

Advanced: Recursive Citation Exploration

For exhaustive citation network analysis:

Level 0: Seed paper (PMID 12345)
├─ Level 1: Forward + backward citations (dispatch 2 subagents)
│  ├─ Find 12 relevant papers
│  └─ Add to papers-reviewed.json
├─ Level 2: For each of 12 papers, explore citations (dispatch 24 subagents)
│  ├─ Find 43 new relevant papers
│  └─ Add to papers-reviewed.json
└─ Level 3: For top 10 papers from Level 2, explore citations
   ├─ Find 28 new relevant papers
   └─ STOP (reaching diminishing returns)

Total: 83 papers discovered through citation network

Stopping criteria:

• Max depth (e.g., 3 levels)
• Max total papers (e.g., 200)
• Diminishing returns (fewer relevant papers per level)
• Time/cost budget

Next Steps After Subagent Review

1. Review consolidated results for quality and consistency
2. Identify gaps - any expected papers missing?
3. Deep dive on highly relevant papers (if not already done)
4. Generate final summary with statistics and key findings
5. Plan next actions - citation exploration? Specific data extraction?

See Also

• evaluating-paper-relevance: Core screening methodology
• traversing-citations: Citation exploration techniques
• finding-open-access-papers: Unpaywall integration for paywalled papers