Parallel Investigation
Coordinate parallel investigation threads to explore multiple hypotheses simultaneously. Most effective for production incidents, performance regressions, or integration failures where the root cause is unclear.
Core Principle
When uncertain, explore multiple paths in parallel. Converge when evidence points to an answer.
Parallel investigation reduces time-to-solution by eliminating serial bottlenecks.
Investigation Structure
Phase 1: Problem Decomposition
Break the problem into independent investigation threads:
Problem: API responses are slow
Investigation Threads:
├── Thread A: Database performance
│ └── Check slow queries, indexes, connection pool
├── Thread B: Application code
│ └── Profile endpoint handlers, check for N+1
├── Thread C: Infrastructure
│ └── Check CPU, memory, network latency
└── Thread D: External services
└── Check third-party API response times
Each thread should be independent (no blocking dependencies), focused (clear scope), and time-boxed.
Phase 2: Thread Assignment
Assign threads with clear ownership:
## Thread A: Database Performance
**Investigator:** [Name/Agent A]
**Duration:** 30 minutes
**Scope:**
- Query execution times
- Index utilization
- Connection pool metrics
**Report Format:** Summary + evidence
Phase 3: Parallel Execution
Each thread follows this pattern:
- Gather evidence specific to thread scope
- Document findings as you go
- Identify if thread is a lead or dead end
- Prepare summary for sync point
Thread Log Template:
## Thread: [Name]
**Start:** [Time]
### Findings
- [Timestamp] [Finding]
### Evidence
- [Log/Metric/Screenshot]
### Preliminary Conclusion
[What this thread suggests about the problem]
Phase 4: Sync Points
Regular convergence to share findings:
Sync Point Agenda:
1. Each thread report (2 min each)
2. Discussion & correlation (5 min)
3. Decision: Continue, Pivot, or Converge (3 min)
Sync Point Decisions:
- Continue: Threads are progressing, maintain parallel execution
- Pivot: Redirect threads based on new evidence
- Converge: One thread found the answer, others join to validate
Phase 5: Convergence
When a thread identifies the likely cause:
- Validate — Other threads verify the finding
- Deep dive — Focused investigation on identified cause
- Document — Compile findings from all threads
Coordination Patterns
Hub and Spoke: One coordinator assigns threads, tracks progress, calls sync points, and makes convergence decisions. Best when one person has the most context.
Peer Network: Equal investigators post findings to a shared channel and self-organize convergence when a pattern emerges. Best when investigators have similar expertise.
Communication Protocol
During Investigation
[Thread A] [Status] Starting query analysis
[Thread B] [Finding] No N+1 patterns in user endpoint
[Thread A] [Finding] Slow query: SELECT * FROM orders WHERE...
[Thread C] [Dead End] CPU and memory within normal
[Thread A] [Hot Lead] Missing index on orders.user_id
At Sync Point
## Thread A Summary
**Status:** Hot Lead
**Key Finding:** Missing index on orders.user_id
**Evidence:** Query taking 3.2s, explain shows full table scan
**Recommendation:** Likely root cause — suggest converge
Decision Framework
| Thread Status | Action |
|---|
| All exploring | Continue parallel |
| One hot lead | Validate lead, others support |
| Multiple leads | Prioritize by evidence strength |
| All dead ends | Reframe problem, new threads |
| Confirmed cause | Converge, begin fix |
Time Management
A typical two-hour investigation:
0:00 Problem decomposition & thread assignment
0:15 Parallel investigation begins
0:45 Sync point #1 → Continue/Pivot/Converge decision
1:30 Sync point #2 (if continuing)
1:35 Final convergence & documentation
Adjust sync point cadence based on incident severity — every 20 minutes for critical outages, every 45 minutes for lower-urgency investigations.
Documentation
Final Report Structure
# Investigation: [Problem]
## Summary
[Brief description and resolution]
## Threads Explored
### Thread A: [Area]
- Investigator: [Name]
- Findings: [Summary]
- Outcome: [Lead / Dead End / Root Cause]
## Root Cause
[Detailed explanation of what was found]
## Evidence
- [Evidence 1]
- [Evidence 2]
## Resolution
[What was done to fix]
## Lessons Learned
- [Learning 1]
Integration with Other Skills
- debugging/root-cause-analysis: Each thread follows RCA principles
- debugging/hypothesis-testing: Threads test specific hypotheses
- handoff-protocols: When passing a thread to another person
anthropics/skills180k
