Hybrid Intelligence

Automatic complexity detection that determines when to use specialized agents.

The Concept

Not every task needs deep analysis. Hybrid Intelligence scores your prompt and chooses the appropriate path:

Score	Path	Time	Agent
0-4	Simple	~2s	None
5-9	Moderate	~5s + optional	User choice
10-19	Complex	~20s	Automatic
20+	Research	~60s+	Multi-agent

How It Works

Step 1: Trigger Detection

Your prompt is analyzed for complexity triggers:

Input: "Add authentication following existing patterns"

Triggers matched:
- "authentication" → Cross-cutting concern (+4)
- "existing patterns" → Pattern detection (+6)

Total Score: 10 (Complex)

Step 2: Path Selection

Based on score:

Score: 10 → Complex path
Action: Spawn Explore Agent automatically

Step 3: Agent Execution

The appropriate agent analyzes your codebase:

Explore Agent running...

Findings:
- Framework: Express.js 4.18
- Existing auth: JWT pattern in payment-service.js
- User model: Has password field already
- Middleware folder: Empty, ready for auth.js

Step 4: Enhanced Response

Agent findings inform the response:

Based on codebase analysis:
- I'll follow the JWT pattern from payment-service.js
- Using existing User model with password field
- Creating auth middleware in middleware/auth.js

Complexity Triggers

Defined in complexity-rules.json

Trigger	Weight	Detection Keywords
Multi-file scope	+5	"all", "every", "across", "throughout"
Architecture questions	+7	"how should", "best approach", "architecture"
Pattern detection	+6	"like", "similar to", "existing pattern"
Feasibility checks	+4	"possible", "can we", "is it feasible"
Refactoring tasks	+5	"refactor", "restructure", "reorganize"
Cross-cutting concerns	+4	"logging", "auth", "validation", "caching"
Implementation planning	+3	"implement", "add feature", "build"

Risk Multipliers

User factors can increase scores:

Factor	Multiplier
Security-critical	1.5x
High-impact	1.3x
Multi-team	1.2x
External dependency	1.1x

Agent Types

Explore Agent

Aspect	Details
Model	Haiku
Time	~30 seconds
Purpose	Codebase exploration, pattern detection

Capabilities:

• Scan project structure
• Find related implementations
• Detect naming conventions
• Identify frameworks and tools

Plan Agent

Aspect	Details
Model	Sonnet
Time	~60 seconds
Purpose	Implementation planning

Capabilities:

• Design implementation strategies
• Consider trade-offs
• Plan step-by-step approach
• Identify dependencies

Security Agent

Aspect	Details
Model	Sonnet
Time	~45 seconds
Purpose	Vulnerability detection

Capabilities:

• OWASP Top 10 analysis
• Input validation review
• Authentication checks
• Authorization analysis

Performance Agent

Aspect	Details
Model	Sonnet
Time	~45 seconds
Purpose	Bottleneck detection

Capabilities:

• N+1 query detection
• Memory usage analysis
• Algorithm complexity review
• Caching opportunities

Pattern Agent

Aspect	Details
Model	Haiku
Time	~30 seconds
Purpose	Convention consistency

Capabilities:

• Naming convention analysis
• Code style consistency
• Architectural pattern detection
• Best practice verification

Caching

Agent results are cached for 10-20x speedup:

Cache Key Components

hash(
  prompt_text +
  relevant_file_hashes +
  git_branch +
  agent_template
)

Cache Settings

Setting	Value
TTL	24 hours
Max size	50 MB
Storage	.claude/cache/agent-results/
Strategy	Content hash

Cache Invalidation

Automatically invalidates when:

• Referenced files change
• Git branch switches
• Manual cache clear

Multi-Agent Verification

For high complexity (15+) or critical operations:

Complexity: 18 (Very High)
Triggering multi-agent verification...

Agent 1 (Explore): Breadth-first analysis
Agent 2 (Pattern): Convention checking
Agent 3 (Security): Vulnerability scan

Consensus Analysis:
- All agree: Use repository pattern
- Disagreement: Error handling approach
  - Explore: Custom error classes
  - Security: Built-in error types
  Recommendation: Custom classes for better audit trail

Configuration

complexity-rules.json

{
  "rules": [
    {
      "name": "multi_file_scope",
      "weight": 5,
      "triggers": ["all", "every", "across"],
      "description": "Task spans multiple files"
    }
  ],
  "thresholds": {
    "simple": { "max": 4 },
    "moderate": { "min": 5, "max": 9 },
    "complex": { "min": 10, "max": 19 },
    "research": { "min": 20 }
  },
  "user_factors": {
    "security_critical": 1.5,
    "high_impact": 1.3
  }
}

agent-templates.json

{
  "templates": {
    "explore_codebase_context": {
      "type": "Explore",
      "model": "haiku",
      "prompt": "Analyze the codebase for..."
    }
  },
  "trigger_mappings": {
    "pattern_detection": "explore_codebase_context",
    "security_concerns": "security_audit_specialist"
  }
}

Best Practices

Trust the System

If complexity detection says you need an agent, let it run. The 20-second investment saves hours of rework.

Don't Fight Moderate

At score 5-9, you're asked if you want agent help. Say yes if:

• You're unsure about existing patterns
• The codebase is unfamiliar
• Implementation quality matters

Override When Needed

Use flags to control behavior:

/prompt-technical --manual Add simple logging
/prompt-technical --agent Add minor validation

Related

• Phase 0 - Prompt perfection before agents
• Predictive Intelligence - Phase 0.15
• Multi-Agent Research - Deep analysis
• Caching - Result persistence