src/agents/README.md

🤖 Cidadão.AI Multi-Agent System

📋 Overview

The Multi-Agent System is the cognitive core of Cidadão.AI, featuring 17 specialized AI agents with Brazilian cultural identities. Each agent embodies specific expertise in transparency analysis, from anomaly detection to policy evaluation, working together through sophisticated coordination patterns and self-reflection mechanisms.

🏗️ Architecture

src/agents/
├── deodoro.py          # Base agent framework & communication protocols  
├── abaporu.py          # Master agent - investigation orchestration
├── zumbi.py            # Investigator - anomaly detection specialist
├── anita.py            # Analyst - pattern analysis expert  
├── tiradentes.py       # Reporter - natural language generation
├── ayrton_senna.py     # Semantic router - intelligent query routing
├── nana.py             # Memory agent - multi-layer memory management
├── machado.py          # Textual analyst - document processing
├── bonifacio.py        # Policy analyst - institutional effectiveness
├── dandara.py          # Social justice - equity monitoring
├── drummond.py         # Communication - multi-channel messaging
├── maria_quiteria.py   # Security auditor - system protection
├── niemeyer.py         # Visualization - data architecture
├── ceuci.py            # ETL specialist - data processing
├── obaluaie.py         # Health monitor - wellness tracking
└── lampiao.py          # Regional analyst - territorial insights

🧠 Agent Coordination Patterns

Master-Agent Hierarchy

# Central coordination with adaptive strategies
MasterAgent (Abaporu)
├── coordinates → InvestigatorAgent (Zumbi)
├── coordinates → AnalystAgent (Anita)  
├── coordinates → ReporterAgent (Tiradentes)
├── coordinates → SemanticRouter (Ayrton Senna)
└── coordinates → ContextMemoryAgent (Nanã)

# Self-reflection loops with quality thresholds
Reflection Loop:
1. Execute investigation
2. Assess quality (threshold: 0.8)
3. If quality < threshold: reflect & adapt
4. Max 3 reflection iterations
5. Return optimized results

Communication Architecture

# Structured message passing between agents
AgentMessage:
- sender: str               # Agent identifier
- recipient: str            # Target agent
- action: str              # Action to perform
- payload: Dict[str, Any]  # Message data
- context: AgentContext    # Shared investigation context
- requires_response: bool  # Synchronous vs async

AgentResponse:
- agent_name: str          # Responding agent
- status: AgentStatus      # Success/failure/in_progress
- result: Any              # Actual result data
- error: Optional[str]     # Error details if failed
- metadata: Dict           # Processing metrics

🎭 Agent Profiles

1. Abaporu - Master Agent (Orchestrator)

Cultural Reference: Abaporu painting by Tarsila do Amaral - symbol of Brazilian Modernism

# Core capabilities
MasterAgent:
- Investigation planning with adaptive strategies
- Agent registry and dependency management  
- Self-reflection with configurable thresholds
- Quality assessment and strategy adaptation
- Comprehensive result synthesis

# Advanced features
- Reflection threshold: 0.8 (configurable)
- Max reflection loops: 3 iterations
- Adaptive investigation strategies based on results
- Agent capability matching and load balancing

Key Methods:

• plan_investigation() - Creates adaptive investigation strategies
• coordinate_agents() - Orchestrates multi-agent workflows
• reflect_on_results() - Self-assessment and strategy adaptation
• synthesize_findings() - Combines results from multiple agents

2. Zumbi - Investigator Agent (Anomaly Detective)

Cultural Reference: Zumbi dos Palmares - freedom fighter and resistance leader

# Anomaly detection capabilities
InvestigatorAgent:
- Price anomalies: 2.5 standard deviation threshold
- Vendor concentration: 70% concentration trigger
- Temporal patterns: Fourier transform analysis
- Duplicate detection: 85% similarity threshold
- Payment irregularities: Statistical outlier detection

# Advanced analytics
- Spectral analysis using FFT for periodic patterns
- Multi-dimensional anomaly scoring
- Machine learning-based pattern recognition
- Cryptographic evidence verification

Anomaly Types:

• PRICE_ANOMALY - Statistical price outliers
• VENDOR_CONCENTRATION - Monopolistic vendor patterns
• TEMPORAL_SUSPICION - Suspicious timing patterns
• DUPLICATE_CONTRACT - Contract similarity detection
• PAYMENT_IRREGULARITY - Payment pattern analysis

3. Anita Garibaldi - Analyst Agent (Pattern Expert)

Cultural Reference: Anita Garibaldi - revolutionary and feminist pioneer

# Pattern analysis capabilities
AnalystAgent:
- Spending trend analysis with linear regression
- Organizational behavior pattern comparison
- Vendor behavior analysis across organizations
- Seasonal pattern detection (end-of-year analysis)
- Cross-spectral analysis between entities
- Efficiency metrics calculation

# Advanced features
- Time series decomposition (trend, seasonal, residual)
- Cross-correlation analysis between organizations
- Spectral density estimation for periodic spending
- Multi-variate regression for complex patterns

Analysis Types:

• SPENDING_TRENDS - Linear regression trend analysis
• VENDOR_PATTERNS - Vendor behavior profiling
• ORGANIZATIONAL_BEHAVIOR - Cross-org comparison
• SEASONAL_ANALYSIS - Seasonal spending patterns
• EFFICIENCY_METRICS - Performance indicators

4. Tiradentes - Reporter Agent (Communication Expert)

Cultural Reference: Tiradentes - independence martyr and symbol of justice

# Report generation capabilities
ReporterAgent:
- Multi-format generation: Markdown, HTML, PDF, JSON
- Audience adaptation: technical, executive, public
- Executive summary creation with key insights
- Risk assessment and prioritization
- Multilingual support: PT-BR, EN-US

# Advanced features
- Template-based report generation
- Natural language explanation of technical findings
- Visualization integration with charts and graphs
- Compliance report formatting for regulatory bodies

Report Formats:

• EXECUTIVE_SUMMARY - High-level findings for executives
• TECHNICAL_REPORT - Detailed analysis for specialists
• PUBLIC_REPORT - Citizen-friendly transparency reports
• COMPLIANCE_REPORT - Regulatory compliance documentation

5. Ayrton Senna - Semantic Router (Query Intelligence)

Cultural Reference: Ayrton Senna - Formula 1 champion symbolizing precision and speed

# Intelligent routing capabilities
SemanticRouter:
- Rule-based routing with regex patterns
- Semantic similarity analysis for complex queries
- Intent detection for conversational flows
- Fallback strategies for ambiguous cases
- Agent capability matching and load balancing

# Routing strategies
1. Rule-based: Fast pattern matching for common queries
2. Semantic: Vector similarity for complex queries  
3. Fallback: Default routing when ambiguous

Query Types:

• INVESTIGATION_QUERY → InvestigatorAgent
• ANALYSIS_QUERY → AnalystAgent
• REPORT_REQUEST → ReporterAgent
• MEMORY_QUERY → ContextMemoryAgent

6. Nanã - Context Memory Agent (Wisdom Keeper)

Cultural Reference: Nanã - Yoruba deity of wisdom and ancestral memory

# Multi-layer memory architecture
ContextMemoryAgent:
- Episodic memory: Investigation results and events
- Semantic memory: General knowledge and patterns
- Conversational memory: Dialog context preservation
- Memory importance scoring and decay management
- Vector-based semantic search with ChromaDB

# Memory layers
Episodic: Specific investigation events and results
Semantic: General patterns and knowledge base  
Conversational: Dialog context and user preferences

Memory Operations:

• store_episodic() - Store investigation results
• retrieve_semantic() - Query knowledge patterns
• maintain_conversation() - Preserve dialog context
• consolidate_memory() - Long-term memory formation

7. Machado de Assis - Textual Analyst (Document Master)

Cultural Reference: Machado de Assis - greatest Brazilian writer and literary genius

# Document processing capabilities
TextualAnalyst:
- Document classification: contracts, laws, decrees
- Named Entity Recognition: organizations, values, dates
- Suspicious clause identification using regex patterns
- Legal compliance checking against frameworks
- Readability assessment (Portuguese-adapted Flesch)
- Transparency scoring based on document clarity

# NLP pipeline
1. Document classification and structure analysis
2. Named entity extraction and relationship mapping
3. Suspicious pattern detection in legal text
4. Compliance validation against regulatory frameworks
5. Readability and transparency scoring

8. José Bonifácio - Policy Analyst (Institutional Architect)

Cultural Reference: José Bonifácio - Patriarch of Independence and institutional architect

# Policy effectiveness evaluation
PolicyAnalyst:
- Efficacy assessment: Did the policy achieve its goals?
- Efficiency evaluation: Resource utilization analysis
- Effectiveness measurement: Impact vs. cost analysis
- Social Return on Investment (SROI) calculation
- Beneficiary impact analysis and coverage assessment
- Sustainability scoring (0-100 scale)

# Evaluation frameworks
- Logic Model: Inputs → Activities → Outputs → Outcomes
- Theory of Change: Causal pathway analysis
- Cost-Benefit Analysis: Economic impact assessment
- Social Impact Measurement: Beneficiary outcome tracking

9. Dandara - Social Justice Agent (Equity Guardian)

Cultural Reference: Dandara dos Palmares - warrior for social justice and equality

# Equity monitoring capabilities
SocialJusticeAgent:
- Gini coefficient calculation for inequality measurement
- Equity violation detection using statistical methods
- Inclusion gap identification across demographics
- Distributive justice assessment
- Intersectional analysis capabilities
- Social vulnerability mapping

# Inequality indices
- Gini Coefficient: Income/resource distribution
- Atkinson Index: Inequality aversion measurement
- Theil Index: Decomposable inequality measure
- Palma Ratio: Top 10% vs. bottom 40% comparison

🔄 Agent Lifecycle & State Management

Agent States

class AgentStatus(Enum):
    IDLE = "idle"                    # Ready for new tasks
    PROCESSING = "processing"        # Currently executing
    REFLECTING = "reflecting"        # Self-assessment phase
    WAITING = "waiting"             # Waiting for dependencies
    COMPLETED = "completed"         # Task finished successfully
    ERROR = "error"                 # Execution failed
    TIMEOUT = "timeout"             # Execution exceeded time limit

State Transitions

# Normal execution flow
IDLE → PROCESSING → COMPLETED
     ↓           ↓
   ERROR    REFLECTING → PROCESSING (adaptive retry)
             ↓
           COMPLETED (after improvement)

# Timeout handling
PROCESSING → TIMEOUT → ERROR (cleanup)

🧪 Self-Reflection Mechanisms

Quality Assessment Framework

class ReflectionMetrics:
    confidence_score: float     # Result confidence (0-1)
    completeness: float        # Investigation thoroughness (0-1)  
    consistency: float         # Internal consistency (0-1)
    novelty: float            # New insights discovered (0-1)
    actionability: float      # Practical usefulness (0-1)

# Reflection triggers
if overall_quality < reflection_threshold:
    reflect_and_improve()

Adaptive Strategies

# Strategy adaptation based on reflection
ReflectionResult:
- quality_issues: List[str]    # Identified problems
- improvement_plan: str        # How to improve
- strategy_adjustments: Dict   # Parameter changes
- confidence_boost: float     # Expected improvement

# Example adaptations
Low confidence → Increase data sampling
Missing patterns → Add analysis techniques  
Incomplete coverage → Expand search criteria

💾 Memory Architecture

Multi-Layer Memory System

# Episodic Memory - Specific events and investigations
EpisodicMemory:
- investigation_results: Dict[str, InvestigationResult]
- agent_interactions: List[AgentMessage]  
- user_queries: List[QueryContext]
- temporal_indexing: Dict[datetime, List[str]]

# Semantic Memory - General knowledge and patterns
SemanticMemory:
- anomaly_patterns: Dict[str, PatternTemplate]
- organization_profiles: Dict[str, OrgProfile]
- vendor_behaviors: Dict[str, VendorProfile]  
- legal_knowledge: Dict[str, LegalConcept]

# Conversational Memory - Dialog context
ConversationalMemory:
- user_preferences: Dict[str, Any]
- conversation_history: List[Message]
- context_stack: List[Context]
- session_metadata: Dict[str, Any]

Memory Operations

# Memory storage with importance weighting
await memory_agent.store_episodic(
    event="investigation_completed",  
    data=investigation_result,
    importance=0.9,  # High importance
    decay_rate=0.1   # Slow decay
)

# Semantic retrieval with vector search
similar_patterns = await memory_agent.retrieve_semantic(
    query_vector=embedding,
    similarity_threshold=0.8,
    max_results=10
)

# Conversational context maintenance
context = await memory_agent.get_conversation_context(
    user_id="user123",
    lookback_messages=20
)

🛡️ Security & Ethics

Agent Security Framework

# Input validation and sanitization
@security_guard
async def process_investigation(query: str) -> InvestigationResult:
    # 1. Input sanitization
    sanitized_query = sanitize_input(query)
    
    # 2. Permission validation  
    validate_permissions(user_context)
    
    # 3. Rate limiting per agent
    await rate_limiter.check_agent_limits(agent_name)
    
    # 4. Audit logging
    await audit_logger.log_agent_action(...)

# Ethics guard - prevents harmful analyses
EthicsGuard:
- prevent_privacy_violations()
- ensure_transparency_goals()  
- validate_public_interest()
- block_discriminatory_analysis()

Audit Trail

# Complete agent action logging
AgentAuditEvent:
- agent_name: str                # Which agent
- action: str                   # What action
- input_data: Dict             # Input parameters (sanitized)
- output_summary: str          # Output summary (no sensitive data)
- success: bool                # Success/failure
- processing_time: float       # Performance metrics
- timestamp: datetime          # When it occurred
- user_context: UserContext    # Who requested it

🧪 Testing Strategy

Agent Testing Framework

# Unit tests for individual agent logic
@pytest.mark.unit
async def test_investigator_price_anomaly_detection():
    agent = InvestigatorAgent()
    data = create_test_contracts_with_price_anomaly()
    
    result = await agent.detect_price_anomalies(data)
    
    assert len(result.anomalies) == 1
    assert result.anomalies[0].type == "PRICE_ANOMALY"
    assert result.anomalies[0].confidence > 0.8

# Integration tests for agent communication
@pytest.mark.integration  
async def test_master_agent_investigation_workflow():
    master = MasterAgent()
    investigator = InvestigatorAgent()
    reporter = ReporterAgent()
    
    # Register agents
    master.register_agent("investigator", investigator)
    master.register_agent("reporter", reporter)
    
    # Execute full workflow
    result = await master.conduct_investigation(
        query="Analyze suspicious contracts",
        agents=["investigator", "reporter"]
    )
    
    assert result.status == "completed"
    assert len(result.findings) > 0
    assert result.report is not None

Mock Agent System

# Mock agents for testing without external dependencies
class MockInvestigatorAgent(InvestigatorAgent):
    async def detect_anomalies(self, data):
        # Return predictable test results
        return create_mock_anomaly_results()

# Test fixtures with realistic data
@pytest.fixture
def sample_investigation_data():
    return {
        "contracts": create_test_contracts(count=1000),
        "vendors": create_test_vendors(count=100),
        "organizations": create_test_organizations(count=50)
    }

📊 Performance Metrics

Agent Performance Monitoring

# Performance metrics per agent
AgentMetrics:
- average_processing_time: float    # Mean execution time
- success_rate: float              # Success percentage
- reflection_frequency: float      # How often reflection occurs
- quality_scores: List[float]      # Historical quality metrics
- memory_usage: float             # Memory consumption
- cache_hit_rate: float           # Cache efficiency

# System-wide metrics
SystemMetrics:
- total_investigations: int        # Total investigations completed
- average_coordination_time: float # Master agent coordination time
- agent_utilization: Dict[str, float] # Per-agent usage
- error_rates: Dict[str, float]   # Per-agent error rates

Scaling Patterns

# Horizontal scaling with agent pools
AgentPool:
- pool_size: int = 5              # Number of agent instances
- load_balancing: str = "round_robin"  # Distribution strategy
- health_checks: bool = True      # Monitor agent health
- auto_scaling: bool = True       # Dynamic scaling based on load

# Vertical scaling with resource limits
ResourceLimits:
- max_memory_mb: int = 1024       # Memory limit per agent
- max_processing_time: int = 300  # Timeout in seconds
- max_concurrent_tasks: int = 10  # Concurrent task limit

🚀 Development & Deployment

Local Development

# Run individual agent tests
pytest tests/unit/agents/test_investigator.py -v

# Run multi-agent integration tests  
pytest tests/integration/agents/ -v

# Performance testing with realistic data
pytest tests/performance/agents/ --benchmark-only

# Memory profiling
pytest tests/agents/ --memray

Agent Configuration

# Environment-specific agent configuration
AgentConfig:
    reflection_threshold: float = 0.8    # Quality threshold
    max_reflection_loops: int = 3        # Max improvement iterations
    memory_retention_days: int = 90      # Memory retention period
    enable_learning: bool = False        # Online learning (experimental)
    parallel_processing: bool = True     # Concurrent agent execution
    
# Per-agent configuration
INVESTIGATOR_CONFIG = {
    "anomaly_threshold": 2.5,           # Standard deviations for anomalies
    "similarity_threshold": 0.85,       # Duplicate detection threshold
    "max_records_per_batch": 10000     # Batch processing size
}

Docker Deployment

# Multi-agent container with resource limits
FROM python:3.11-slim

# Install agent dependencies
COPY requirements/agents.txt .
RUN pip install -r agents.txt

# Copy agent source code
COPY src/agents/ /app/agents/

# Resource limits for agent container
ENV MEMORY_LIMIT=2048MB
ENV CPU_LIMIT=2.0
ENV MAX_AGENTS=10

# Health check for agent system
HEALTHCHECK --interval=30s --timeout=10s \
  CMD python -c "from src.agents import health_check; health_check()"

CMD ["python", "-m", "src.agents.orchestrator"]

🔮 Future Enhancements

Planned Features

• Federated Learning: Agents learn from distributed investigations
• Dynamic Agent Creation: Generate specialized agents for new domains
• Cross-Language Support: Multi-language document analysis
• Real-time Collaboration: Simultaneous multi-agent processing
• Explainable AI: Enhanced transparency in agent decision-making

Research Areas

• Agent Personality Development: More sophisticated cultural personas
• Emotional Intelligence: Agents that understand social context
• Creative Problem Solving: Novel approach generation for complex problems
• Meta-Learning: Agents that improve their learning strategies

---

This multi-agent system represents a unique approach to transparency analysis, combining cutting-edge AI with Brazilian cultural identity to create agents that are both technically sophisticated and culturally meaningful. Each agent contributes specialized expertise while working together through advanced coordination patterns to democratize access to government transparency analysis.