Parallel Research Orchestration

A systematic methodology for conducting thorough research using hierarchical multi-agent coordination.

When to Use

Deep investigation of complex topics
Research requiring multiple perspectives
Analysis with potential for conflicting viewpoints
Exploratory work in unfamiliar domains
Any research that benefits from devil's advocate review

Agent Hierarchy

1. Lead Agent (Coordinator)

Decomposes research question into threads
Assigns threads to sub-agents
Monitors progress and adjusts strategy
Coordinates synthesis

2. Sub-Agents (Specialists)

Execute specific research threads
Focus deeply on assigned topic
Report findings with confidence levels
Identify gaps and uncertainties

3. Critical Review Agent (Devil's Advocate)

Challenges findings from sub-agents
Identifies weaknesses in reasoning
Proposes alternative interpretations
Stress-tests conclusions

4. Synthesis Agent (Integrator)

Combines findings across threads
Resolves conflicts between sources
Produces coherent narrative
Highlights remaining uncertainties

Phase 1: Anticipatory Decomposition

Lead Agent Responsibilities:

Analyze Research Question
- Identify core question and sub-questions
- Map dependencies between research threads
- Anticipate potential failures and alternative remedies
Decompose Into Parallel Threads
- Create 3-6 independent research threads
- Each thread: clear objective, search strategy, success criteria
- Ensure threads cover different perspectives (not redundant)
- Define explicit handoff protocols

Launch Sub-Agents

Launch 4 research subagents in parallel:
- Thread 1: [Specific focus and search strategy]
- Thread 2: [Specific focus and search strategy]
- Thread 3: [Specific focus and search strategy]
- Thread 4: [Specific focus and search strategy]

Each agent should return only:
- Key findings
- Evidence quality assessment
- Confidence score
- Conflicting information found

Phase 2: Parallel Research Execution

Research Sub-Agent Instructions:

Search Strategy
- Execute assigned searches
- Use progressive disclosure (don't front-load context)
- Cross-reference multiple sources for balance
- Track evidence quality
Validation Requirements
- Verify claims against original sources
- Note contradictions or conflicts found
- Assign confidence scores (high/medium/low)
- Flag assumptions or gaps

Return Format

Thread [N] Findings:

KEY FINDINGS:
- [Finding 1] (Confidence: high/medium/low)
- [Finding 2] (Confidence: high/medium/low)

EVIDENCE QUALITY:
- [Source type, credibility, date]

CONFLICTS DETECTED:
- [Any contradictions between sources]

GAPS/LIMITATIONS:
- [What wasn't found or remains unclear]

Phase 3: Critical Review (Devil's Advocate)

Role Assignment: "You are a systematic skeptic. Your role is to identify risks, edge cases, failure modes, logical fallacies, and vulnerabilities. Focus on disagreement and counterarguments, not confirmation."

Three-Fold Review:

Anticipatory Critique
- What could be wrong with these findings?
- What alternative interpretations exist?
- What evidence is missing or weak?
Finding-by-Finding Challenge
- For each finding: "What if this is wrong?"
- Identify logical fallacies or reasoning gaps
- Check for confirmation bias
Strategic Refinement
- What should agents have done differently?
- Which low-confidence findings should be rejected?
- What additional research is needed?

Output Requirements:

CRITICAL REVIEW REPORT:

STRONG FINDINGS (accept):
- [Findings that withstand scrutiny]

WEAK FINDINGS (reject or flag):
- [Findings with logical flaws, weak evidence]

IDENTIFIED RISKS:
- [Edge cases, failure modes]

CONFLICTS REQUIRING RESOLUTION:
- [Contradictions between threads]

ADDITIONAL RESEARCH NEEDED:
- [Gaps requiring follow-up]

Phase 4: Conflict Resolution & Synthesis

Conflict Resolution Framework:

Debate Pattern
- For each conflict, have research agents defend findings
- Evaluate on: evidence quality, logical consistency, edge case coverage
- Judge agent makes final decision
Voting with Confidence Weighting
- Weight findings by confidence scores
- Prioritize primary over secondary sources
- Prioritize recent over outdated (when relevant)
- Require minimum confidence threshold
Cross-Referencing Validation
- Verify final synthesis against original sources
- Ensure balanced perspective
- Flag remaining uncertainties

Synthesis Output Format:

RESEARCH SYNTHESIS REPORT

EXECUTIVE SUMMARY:
[2-3 paragraph synthesis answering original question]

VALIDATED FINDINGS:
1. [Finding] (Confidence: X, Sources: Y)
2. [Finding] (Confidence: X, Sources: Y)

CONFLICTS RESOLVED:
- [How contradictions were resolved]

REMAINING UNCERTAINTIES:
- [What remains unclear]

RECOMMENDATIONS:
- [Actionable recommendations]

SOURCES:
- [Complete source list with hyperlinks]

Quality Assurance Principles

Parallel Validation: Multiple verification paths catch different errors
End-State Evaluation: Focus on correct final synthesis, not process
Separation of Generation and Evaluation: Research and critique are separate
Transparency: Full disclosure of methodology and sources
Human Escalation: For irresolvable conflicts, escalate to user

Claude Code Specific Optimizations

Use Built-in Subagents: Leverage Plan Subagent for orchestration, Explore Subagent for codebase research
Parallel Execution: Always execute independent threads in parallel
Context Preservation: Main thread maintains context; subagents use isolated windows
Token Awareness: This methodology uses significant tokens (4+ parallel agents)

Common Pitfalls to Avoid

Sequential Research: Don't run threads one-by-one; parallelize
Weak Devil's Advocate: Enforce systematic skepticism
Premature Synthesis: Don't synthesize before critical review
Context Bleeding: Keep sub-agent contexts isolated
Unchallenged Conflicts: Require confidence thresholds and debate
Missing Transparency: Always include source attribution

Success Metrics

A successful research orchestration produces:

High-confidence findings (backed by multiple sources)
Acknowledged and resolved conflicts
Transparent limitations and uncertainties
Actionable synthesis (not just information dump)
Complete source attribution
Evidence of critical review (not confirmation bias)

parallel-research