Multi agent researcher

Multi-Agent Research Coordinator

Purpose

Transform complex research questions into comprehensive reports by:

1. Decomposing broad topics into 2-4 focused subtopics
2. Spawning specialized researcher agents in parallel
3. Synthesizing findings into cohesive final report
4. Saving structured outputs for reference

When to Use

Auto-invoke when user asks:

• Search/Discovery: "Search what is [topic]", "Find information about [subject]", "Look up [technology]", "Discover [patterns]"
• Investigation: "Research [topic]", "Investigate [subject]", "Analyze [phenomenon]", "Study [field]", "Explore [domain]"
• Collection: "Gather information about [subject]", "Collect data on [topic]", "Compile resources for [area]"
• Learning: "Learn about [subject]", "Tell me about [topic]", "Dig into [technology]", "Delve into [concept]"
• Contextual: "What are the latest developments in [field]?", "Comprehensive analysis of [topic]", "Deep dive into [subject]", "State of the art in [domain]", "Best practices for [area]"

Do NOT invoke for:

• Simple factual questions ("What is the capital of France?")
• Decision evaluation ("Should I use X or Y?")
• Code-related tasks ("Debug this function", "Write a script")

Orchestration Workflow

Phase 1: Query Analysis & Decomposition

Step 1.1: Understand the Research Question Analyze user's query to identify core topic, scope, and intent.

Step 1.2: Decompose into Subtopics Break topic into 2-4 focused subtopics that are:

• Mutually exclusive (minimal overlap)
• Collectively exhaustive (cover whole topic)
• Independently researchable
• Together provide comprehensive coverage

Decomposition Patterns:

Temporal: Past → Current → Future Categorical: Category 1, 2, 3 Stakeholder: Technical → Business → Policy → User Problem-Solution: Problem → Solutions → Gaps → Future Geographic: Region A → Region B → Comparison

Step 1.3: Create Research Plan Use TodoWrite to track:

- [ ] Decompose query into subtopics
- [ ] Spawn researcher 1: [subtopic]
- [ ] Spawn researcher 2: [subtopic]
- [ ] Spawn researcher 3: [subtopic]
- [ ] Synthesize findings
- [ ] Save final report

---

Phase 2: Parallel Research Execution

Step 2.1: Spawn Researcher Agents in Parallel

For each subtopic, create a Task tool call with:

subagent_type: "researcher"
description: "Research {subtopic name}"
prompt: "Research the following subtopic in depth:

**Subtopic**: {Subtopic name}
**Context**: Part of research on '{original topic}'
**Focus**: {Specific guidance}

Conduct thorough web research, gather authoritative sources, extract key findings, and save results to files/research_notes/{subtopic-slug}.md"

Critical: Spawn ALL researchers in parallel (multiple Task calls in same message), not sequentially.

Step 2.2: Monitor Completion Update TodoWrite as researchers complete.

Step 2.3: Verify All Complete Use Glob to confirm all files exist: files/research_notes/*.md

---

Phase 3: Synthesis & Report Generation

⚠️ CRITICAL: ARCHITECTURAL ENFORCEMENT ACTIVE ⚠️

YOU DO NOT HAVE WRITE TOOL ACCESS when this skill is active. The allowed-tools frontmatter explicitly EXCLUDES the Write tool to enforce proper workflow delegation.

YOU CANNOT:

• ❌ Write synthesis reports yourself
• ❌ Create files in files/reports/ directory
• ❌ Bypass the report-writer agent

YOU MUST:

• ✅ Spawn report-writer agent via Task tool
• ✅ Delegate all synthesis and report writing to the agent
• ✅ Read the completed report and deliver to user

---

Step 3.1: Verify Research Completion

1. Use Glob to confirm all research notes exist: files/research_notes/*.md
2. Verify count matches number of spawned researchers
3. If any missing: investigate and complete before synthesis

Step 3.2: Spawn Report-Writer Agent (MANDATORY)

This is the ONLY synthesis approach - there is no "Option A" or "Option B". You MUST use the report-writer agent because you lack Write tool permissions.

Task:
subagent_type: "report-writer"
description: "Synthesize research findings into comprehensive report"
prompt: "Synthesize research into comprehensive report:

**Original Question**: {user query}
**Subtopics Researched**: {list all subtopics}
**Notes Location**: files/research_notes/

## Your Tasks:
1. Read ALL research notes from files/research_notes/
2. Identify themes, patterns, and contradictions across notes
3. Synthesize findings into cohesive narrative
4. Cite sources from research notes
5. Add cross-cutting insights beyond individual notes
6. Save comprehensive report to files/reports/{topic-slug}_{timestamp}.md

## Report Structure:
- Executive Summary
- Key Findings (with evidence from research notes)
- Detailed Analysis by subtopic
- Cross-Cutting Themes
- Contradictions and Debates
- Gaps and Limitations
- Source Bibliography

Use the timestamp format: $(date +\"%Y%m%d-%H%M%S\") for the filename."

Step 3.3: Monitor Agent Completion

After spawning report-writer agent, wait for completion. The agent will:

• Read all research notes
• Synthesize findings
• Write comprehensive report to files/reports/
• Return completion message with file path

---

Phase 4: Deliver Results

Step 4.1: Create User Summary

# Research Complete: {Topic}

Comprehensive research completed with {N} specialized researchers.

## Key Findings
1. {Finding 1}
2. {Finding 2}
3. {Finding 3}

## Research Scope
{N} subtopics investigated:
- {Subtopic 1}
- {Subtopic 2}
- {Subtopic 3}

## Files Generated
**Research Notes**: `files/research_notes/`
- {file1}.md
- {file2}.md
- {file3}.md

**Final Report**: `files/reports/{filename}.md`

## Next Steps
{Optional suggestions}

Step 4.2: Update TodoWrite Mark all items complete.

---

Best Practices

Good Decomposition

✅ 2-4 subtopics (sweet spot: 3) ✅ Distinct but related ✅ Comprehensive coverage ✅ Independently researchable

❌ Too many (>5) ❌ Too few (1) ❌ Significant overlap ❌ Too narrow or too broad

Parallel Execution

• Always spawn researchers simultaneously
• Never sequential unless dependent
• Provide context to each researcher
• Reasonable scope (10-15 min each)

Synthesis Quality

• Read ALL notes
• Find connections across subtopics
• Note contradictions explicitly
• Cite sources
• Add insights beyond individual notes

---

Error Handling

Researcher Fails: Try replacement, proceed with others, note gap No Results Found: Accept partial, note limitation Contradictory Findings: Document all perspectives explicitly Unclear Query: Ask clarifying questions first

---

Examples

Query: "Research quantum error correction" Decomposition:

1. Theoretical foundations
2. Hardware implementations
3. Commercial viability Researchers: 3 parallel Synthesis: report-writer agent (MANDATORY)

Query: "Investigate cryptocurrency market 2025" Decomposition:

1. Market metrics & players
2. Regulatory landscape
3. Technology evolution
4. Institutional trends Researchers: 4 parallel Synthesis: report-writer agent (MANDATORY)

---

Remember: Quality depends on good decomposition, thorough researchers, insightful synthesis, and clear user communication.