Multi Agent Parallel Investigation

Multi-Agent Parallel Investigation

Overview

This skill provides a systematic approach to answering complex architecture questions by decomposing them into 4-6 parallel investigations, each with a specialized perspective. Each agent investigates independently, then findings are synthesized into a phased decision framework with concrete recommendations.

Core Pattern: Spawn multiple Task tool agents in parallel → Each investigates specific dimension → Synthesize findings → Deliver phased implementation plan

When to Use This Skill

Use this skill when you encounter:

1. Complex Architecture Decisions with multiple unknowns requiring investigation from different angles (API capabilities, performance, documentation quality, feasibility)
2. Trade-off Analysis where multiple solutions exist with competing priorities (complexity vs performance, cost vs developer UX)
3. Technology Evaluation when assessing if a platform/tool/library meets requirements across multiple dimensions
4. Data Platform Questions in crypto/trading domains with concerns about availability, correctness, observability, and maintainability
5. Implementation Planning when you need phased rollout with clear decision gates based on empirical findings

Typical Question Patterns:

• "How should we distribute our Parquet database to external users?"
• "What's the best approach for querying remote data without downloads?"
• "Should we build a CLI tool, improve documentation, or create an API?"
• "How can we validate data quality across multiple sources?"

Workflow

Step 1: Decompose Question into Investigation Dimensions

Identify 4-6 specialist perspectives needed to answer the question comprehensively:

Common Agent Roles (select 4-6 based on question):

• API Capabilities Analyst - Platform API evaluation, endpoint discovery
• Performance Analyst - Benchmarking, latency measurements, optimization
• Documentation Analyst - Quality assessment, gap identification, UX evaluation
• Feasibility Engineer - Prototyping, proof-of-concept, effort estimation
• Security Analyst - Threat modeling, credential management, compliance
• Cost Analyst - Pricing, rate limits, resource consumption
• Integration Specialist - Third-party service integration, compatibility
• Data Quality Analyst - Schema validation, coverage assessment, correctness

See references/agent-templates.md for detailed role templates.

Step 2: Spawn Agents in Parallel

Use single message with multiple Task tool calls to maximize performance:

I'll spawn 4 parallel agents to investigate this question from different perspectives:

[Task tool call 1: API Capabilities Agent with prompt from agent-templates.md]
[Task tool call 2: Performance Analyst Agent with prompt from agent-templates.md]
[Task tool call 3: Documentation Analyst Agent with prompt from agent-templates.md]
[Task tool call 4: Feasibility Engineer Agent with prompt from agent-templates.md]

Key Requirements:

• Each agent gets ROLE, OBJECTIVE, CONTEXT, DYNAMIC WRITETODO APPROACH, INVESTIGATION QUESTIONS, DELIVERABLES, WORKSPACE
• Agents work independently (no inter-agent communication)
• Each agent uses /tmp/{role-slug}/ workspace for artifacts
• Each agent reports structured findings back to main context

Step 3: Dynamic WriteTodo Within Each Agent

CRITICAL: Agents use emergent task creation, NOT pre-planned task lists.

Pattern:

1. Agent creates ONE initial writeTodo (e.g., "Search for API documentation")
2. Agent completes task → Analyzes findings → Creates NEXT writeTodo based on discoveries
3. Agent marks completed → Executes next → Repeats until investigation complete
4. Let writeTodos emerge naturally from findings

Example Flow (API Capabilities Agent):

writeTodo 1: "Search for GitHub REST API documentation for release assets"
→ Discovery: Found API endpoints for listing release assets
→ writeTodo 2: "Test HTTP range request support on release asset URL"
→ Discovery: Range requests NOT supported
→ writeTodo 3: "Investigate CDN proxy alternatives (jsDelivr, Cloudflare)"
→ Discovery: jsDelivr supports range requests
→ Investigation complete

Step 4: Wait for Agent Reports

Each agent returns structured report with:

• Findings Summary (2-3 sentences)
• Data/Measurements (performance numbers, API endpoints, examples)
• Confidence Level (HIGH/MEDIUM/LOW with %)
• Recommendation (specific action with justification)

Step 5: Synthesize Findings into Decision Framework

Use patterns from references/synthesis-patterns.md:

Choose synthesis pattern based on agent findings:

• Consensus Building - When agents agree on direction, use voting matrix to validate
• Trade-off Matrix - When agents have competing priorities, score solutions across dimensions
• Risk-Based Synthesis - When findings reveal risks, prioritize by likelihood × impact
• 80/20 Synthesis - When one solution covers most use cases, identify highest ROI option
• Confidence Aggregation - When confidence varies, defer low-confidence phases
• Phased Decision Framework - DEFAULT pattern for most investigations

Example Synthesis (Phased Decision Framework):

## Summary of Parallel Investigations

**Agent 1 (API Capabilities)**: GitHub Releases API supports listing assets but NOT range requests. jsDelivr CDN proxy enables range requests with 95% reliability.

**Agent 2 (Performance)**: DuckDB httpfs queries complete in 2.8s (cold start), bandwidth efficiency 97% vs full download. Range requests work via jsDelivr.

**Agent 3 (Documentation)**: Current README rated 7/10. Missing Quick Start, Prerequisites unclear, no copy-paste examples. Gap: remote query workflow.

**Agent 4 (Feasibility)**: CLI tool buildable in 4-6 hours but adds complexity. Documentation improvements solve 80% of use cases in 2 hours.

## Decision Framework

### Phase 1: Documentation Quick Start (Priority: HIGH)
**Objective**: Enable developers to query remote Parquet in <60 seconds
**Based on**: Agent 3 (documentation gaps) + Agent 2 (proven performance)
**Effort**: 2 hours
**Impact**: Solves 80% of use cases (Agent 4 finding)
**Recommendation**: Add Quick Start section with DuckDB httpfs example, jsDelivr URL pattern, prerequisites

### Phase 2: Performance Optimization (Priority: MEDIUM)
**Objective**: Document query optimization patterns (column pruning, filtering)
**Based on**: Agent 2 (performance benchmarks show 10x speedup with WHERE clauses)
**Effort**: 1 hour
**Impact**: Reduces query time from 2.8s → 0.3s for filtered queries
**Decision Criteria**: Proceed after Phase 1 validates user adoption

### Phase 3: CLI Tool (Priority: LOW, Optional)
**Objective**: Standalone tool for non-Python users
**Based on**: Agent 4 (feasibility prototype)
**Effort**: 4-6 hours
**Impact**: Serves remaining 20% of use cases
**Decision Criteria**: Only proceed if user feedback shows demand after Phase 1+2

## Total Effort Estimate
- **Phase 1**: 2 hours (HIGH confidence)
- **Phase 2**: 1 hour (MEDIUM confidence)
- **Phase 3**: 4-6 hours (LOW confidence, optional)
- **Total**: 3-9 hours depending on user feedback

Step 6: Deliver Actionable Recommendations

Final output should include:

1. Summary - What was investigated, what agents found
2. Decision Framework - Phased implementation plan with priorities
3. Success Criteria - Measurable outcomes (from synthesis-patterns.md → Success Criteria Synthesis)
4. Validation Plan - How to test recommendations
5. Next Steps - Immediate action items

Using Bundled Resources

references/agent-templates.md

Contains 10 example agent prompts with complete structure:

• ROLE, OBJECTIVE, CONTEXT
• DYNAMIC WRITETODO APPROACH
• INVESTIGATION QUESTIONS (5-7 specific questions)
• DELIVERABLES (structured output format)
• WORKSPACE (temp directory for artifacts)

Usage: Copy relevant template → Customize CONTEXT and QUESTIONS for your specific investigation → Use as Task tool prompt

references/synthesis-patterns.md

Contains 8 frameworks for synthesizing agent findings:

1. Phased Decision Framework - Default pattern, structures findings into HIGH/MEDIUM/LOW priority phases
2. Consensus-Building Pattern - Voting matrix for conflicting findings
3. Trade-off Matrix - Score solutions across dimensions (Complexity, Cost, Performance, UX)
4. Risk-Based Synthesis - Prioritize by risk mitigation (P0/P1/P2)
5. Confidence Level Aggregation - Defer low-confidence phases
6. 80/20 Synthesis - Identify highest ROI solution
7. Integration Strategy - Structure complementary solutions
8. Success Criteria Synthesis - Measurable outcomes from agent findings

Usage: After agents report findings, select appropriate synthesis pattern → Fill in template with agent data → Present decision framework

Domain Context: Crypto/Trading Data Platforms

This skill is optimized for questions about:

• Data Distribution - How to serve historical OHLCV data, orderbook snapshots, trade ticks
• Query Performance - Remote vs local access, bandwidth optimization, latency requirements
• Documentation Quality - Developer onboarding friction, example coverage, troubleshooting guides
• API Design - REST endpoints, WebSocket streams, bulk download vs query endpoints
• Storage Technologies - Parquet, DuckDB, CSV/JSON, compression formats
• Infrastructure Decisions - GitHub Releases, S3, CDN proxies, self-hosted APIs
• SLO Dimensions - Availability, Correctness, Observability, Maintainability (NOT speed/performance/security)

Example Questions from Domain:

• "Should we use DuckDB httpfs for remote Parquet queries or build a REST API?"
• "How do we balance query latency vs bandwidth efficiency for 20MB Parquet files?"
• "Is documentation sufficient for developers to query our database in <60 seconds?"
• "What's the feasibility of a CLI tool vs extending existing tooling?"

Tips for Success

1. Parallel Execution: Always spawn agents in single message with multiple Task calls (NOT sequential)
2. Role Specialization: Each agent should have narrow, distinct focus (avoid overlap)
3. Dynamic WriteTodos: Agents discover next steps based on findings, not pre-planned lists
4. Empirical Evidence: Agents should measure/test/validate, not speculate
5. Confidence Levels: Agents report HIGH/MEDIUM/LOW confidence with percentages
6. Synthesis Pattern Selection: Choose pattern based on agent findings (consensus vs conflict vs risk)
7. Phased Implementation: Default to 3 phases (HIGH/MEDIUM/LOW priority) with decision gates
8. Measurable Success: Define concrete validation criteria (timing, percentages, user counts)

Common Pitfalls to Avoid

1. Too Many Agents (>6) - Synthesis becomes unwieldy, prefer 4-6 focused roles
2. Pre-planned WriteTodos - Defeats purpose of emergent investigation, let findings guide tasks
3. Sequential Agent Execution - Wastes time, always spawn in parallel unless dependencies exist
4. Vague Investigation Questions - Each agent needs 5-7 specific, answerable questions
5. No Workspace - Agents need /tmp/{role}/ directories for artifacts/tests/prototypes
6. Ignoring Confidence Levels - Low-confidence findings should trigger Phase 3 (optional), not Phase 1
7. Skipping Synthesis - Raw agent reports are not actionable, must synthesize into decision framework
8. No Validation Plan - Recommendations need measurable success criteria and test approach