Multi-Expert Agent

The Multi-Expert Agent is an orchestrator that analyzes complex questions by deploying 3-5 specialized domain experts in parallel, then synthesizes their findings into a comprehensive, confidence-scored analysis.

Capabilities

• Spawns multiple specialized domain experts in parallel for comprehensive analysis
• Detects relevant domains from questions (security, API, UI, database, testing, performance, CI, DevOps)
• Synthesizes expert findings with confidence scoring (High/Medium/Low)
• Identifies agreements between experts (high confidence) and disagreements (needs review)
• Extracts unique insights from individual experts
• Provides prioritized recommended actions

When to Use

Use this agent when:

• Complex questions spanning multiple domains - "Is our authentication implementation secure?" needs security + API + testing + database perspectives
• Security reviews - Analyze code from multiple angles (security, architecture, testing)
• Architecture reviews - Get feedback from domain experts on design decisions
• PR reviews - Multi-perspective code review for confidence in correctness
• End-to-end flow analysis - Trace requests through multiple layers with expert validation
• Debugging hard problems - Multiple possible causes that need expert analysis
• "Best practice" questions - Need validation across multiple domains

How It Works

The agent follows a structured workflow:

CopyUSER QUESTION
      ↓
┌─────────────────┐
│ 1. ANALYZE      │ → Identify 3-5 relevant domains
│ 2. DEPLOY       │ → Spawn experts IN PARALLEL
│ 3. COLLECT      │ → Wait for all results
│ 4. SYNTHESIZE   │ → Combine with confidence scoring
│ 5. REPORT       │ → High-confidence answer
└─────────────────┘

Step 1: Domain Analysis

The agent analyzes the question for domain keywords and selects 3-5 experts:

• 1 PRIMARY expert - Most relevant domain
• 2-4 SUPPORTING experts - Provide additional perspectives

Step 2: Deploy in Parallel

All experts are deployed simultaneously using background execution for efficiency:

CopyTask(
  description: "Security analysis",
  prompt: "FIRST: Read packages/cli/src/core/experts/security/expertise.yaml\n\nQUESTION: ...",
  subagent_type: "agileflow-security",
  run_in_background: true
)

Step 3: Collect Results

Uses TaskOutput with block: true to wait for all expert responses.

Step 4: Synthesize with Confidence Scoring

Analyzes all responses and groups findings:

• AGREEMENT (2+ experts) → High confidence
• UNIQUE INSIGHT (1 expert) → Notable finding
• DISAGREEMENT → Flagged for human review

Step 5: Report

Outputs structured analysis with sections:

• Key Findings (High Confidence) - Agreements between 2+ experts
• Unique Insights - Notable findings from single experts
• Disagreements (Needs Review) - Conflicting opinions
• Recommended Actions - Prioritized next steps

Confidence Scoring

Domain Detection

The agent maps question keywords to domain experts:

Example

User Question: "Is our authentication implementation secure?"

Agent Response:

CopyAnalyzing for relevant domains...
 
Detected domains:
- Security (primary) - auth security analysis
- API (supporting) - implementation review
- Testing (supporting) - coverage verification
- Database (supporting) - credential storage
 
Deploying 4 experts in parallel...
 
## Multi-Expert Analysis: Authentication Security
 
**Confidence**: High (4 experts, strong agreement)
 
### Key Findings (High Confidence)
1. JWT tokens never expire - CRITICAL - Agreed by: Security, API
2. Auth middleware inconsistent - HIGH - Agreed by: Security, API
 
### Unique Insights
- Database: Password hashing is correct (bcrypt with proper salt rounds)
- Testing: Only 40% coverage on auth routes, missing integration tests
 
### Disagreements (Needs Review)
None - all experts aligned on security stance
 
### Recommended Actions
1. Add JWT expiration and refresh token rotation (CRITICAL)
2. Audit all routes for consistent middleware application (HIGH)
3. Increase test coverage to 80% for auth routes (MEDIUM)
4. Add integration tests for auth flow (MEDIUM)

Key Behaviors

• ALWAYS deploys experts in parallel - Never sequentially (efficiency)
• ALWAYS deploys ALL experts in a SINGLE message - Batch Task calls together
• NEVER gives final answer without synthesizing ALL expert responses - All perspectives matter
• ALWAYS includes expertise reading in expert prompts - "FIRST: Read packages/cli/src/core/experts/{domain}/expertise.yaml"
• Deploys 3-5 experts total - Not fewer, not more (optimal synthesis)
• ALWAYS uses TaskOutput with block: true - To collect all results

Tools Available

This agent has access to:

• Read - Access agent expertise files
• Write - Create analysis documents
• Edit - Update existing analyses
• Bash - Run verification commands
• Glob - Search for relevant files
• Grep - Search code and documentation
• Task - Deploy expert agents in parallel
• TaskOutput - Collect expert results

Model Configuration

• Model: Claude Opus 4.5 (Sonnet)
• Reasoning: Full complex analysis with multiple perspectives

Related Agents

• database - Database schema and query expert
• api - REST API endpoint expert
• ui - Frontend component expert
• testing - Test coverage and strategy expert
• security - Security and authentication expert
• performance - Performance optimization expert
• ci - CI/CD pipeline expert
• devops - Infrastructure and deployment expert

When NOT to Use

• Simple single-domain questions - Use direct domain expert instead
• Questions requiring decision-making authority - Multi-Expert provides analysis only
• Sensitive information - Document experts involved before deploying