Root Cause Analysis

Dispatch 3-7 specialized RCA agents in parallel to investigate a problem from different angles, then synthesize findings into a comprehensive report.

Workflow

1. Parse and Size the Swarm

Given a problem_statement:

• Infer agent count from problem complexity (breadth of domains, surface area, symptoms)
• Set agents_count = clamp(inferred_agents, 3, 7) - always use at least 3 agents
• Select focus areas from the taxonomy below - ensure complementary coverage
• Output chosen count, focus areas, and rationale to console

2. Focus Area Taxonomy

Select areas based on problem characteristics. Mix layers and concerns as needed.

Tech Layers:

• UI/Presentation - Frontend components, rendering, user interaction
• API/Backend - Endpoints, request handling, response formation
• Data/Database - Queries, schemas, data integrity, persistence
• Config/Infrastructure - Environment, deployment, external services
• Dependencies - Third-party libs, version conflicts, compatibility

Concerns (Cross-cutting):

• State Management - Application state, caching, synchronization
• Business Logic - Domain rules, validation, workflows
• Integration - Service boundaries, contracts, data transformation
• Performance - Latency, throughput, resource utilization
• Security - Auth, authz, data protection, input validation

3. Dispatch Agents (Parallel)

Use the Task tool with subagent_type="rca" for each agent. Include these headers in each prompt:

<FOCUS_AREA>[Selected Area from Taxonomy]</FOCUS_AREA>

<PROBLEM_STATEMENT>
[Problem statement verbatim]
</PROBLEM_STATEMENT>

Ensure focus areas are complementary - cover different layers/concerns to maximize investigative breadth.

4. Collect and Synthesize

After all agents return their <rca_findings> XML blocks:

• Parse - Extract <evidence>, <observations>, and <layer_status> from each agent's XML
• Collect - Gather all evidence items into a unified view, grouped by type
• Cross-Reference - Look for:
  • Multiple agents pointing to same location/component (convergence)
  • Contradictory observations (indicates complexity)
  • Cleared layers (narrows the search)
• Weigh - Observations with confidence="high" from multiple agents carry more weight
• Check Status - Which layers are suspect vs cleared vs inconclusive?
• Determine Verdict:
  • Root cause identified (high confidence) - multiple agents converge
  • Root cause suspected (medium confidence) - single agent found it, others cleared
  • Inconclusive (low confidence) - need more investigation

5. Render Report

Use the template in references/report-template.md.

Resources

• references/agent-prompt.md - The RCA agent system prompt for Task tool dispatches
• references/report-template.md - The final report template structure