A production-ready autonomous AI system for Enhanced Due Diligence (EDD) investigations using graph-based workflow orchestration with LangGraph and multi-model AI architecture (Claude Sonnet 4.5 + GPT-4o).
DeepAgents is an autonomous research agent that conducts comprehensive Enhanced Due Diligence (EDD) investigations on individuals and entities. It orchestrates multiple AI models through a graph-based workflow to automatically gather intelligence, analyze risks, map relationships, and generate compliance-ready reports.
The system implements graph-based workflow orchestration using LangGraph to coordinate multiple specialized AI models in a stateful research pipeline:
-
Iterative Research Cycle
- Generates strategic search queries based on reflection analysis
- Executes parallel web searches to gather information
- Analyzes findings and extracts insights
- Decides whether to continue searching or finalize report
- Loops until max depth reached, reflection recommends stopping, or stagnation detected
-
Multi-Model Architecture
- Claude Sonnet 4.5: Strategic analysis, query generation, reflection (fast with simple schemas)
- GPT-4o/mini: Web search, entity extraction, connection mapping, report synthesis (reliable structured output)
-
Entity & Risk Intelligence
- Automatically discovers and deduplicates entities (persons, organizations, events)
- Builds relationship graphs with pattern detection
- Identifies risk indicators across multiple categories
- Flags suspicious connections and conflicts of interest
-
Comprehensive Reporting
- Generates structured due diligence reports
- Risk assessment with severity levels (CRITICAL, HIGH, MEDIUM, LOW)
- Complete source attribution and audit trail
- Actionable recommendations
Autonomous Operation
- No human intervention required during research
- Self-improving search through reflection-driven strategy
- Intelligent termination based on progress assessment
Entity Tracking
- LLM-based entity discovery and deduplication
- Handles aliases, variations, and name changes
- Builds comprehensive entity relationship graphs
- Pattern detection and suspicious connection flagging
Risk Assessment
- Multi-category risk detection (fraud, corruption, sanctions, PEP, regulatory, etc.)
- Severity-labeled red flags with evidence
- Source credibility evaluation
- Gap analysis and research limitations
Production-Ready
- Complete audit trail (JSONL format) for compliance
- Type-safe state management with immutable updates
- Configurable models and parameters (YAML-based)
- Dual logging (audit + operational) with optional LangFuse integration
- Error handling and recovery at all levels
┌─────────────────────────────────────────────────────────┐
│ User Interface │
│ CLI / Python API / REST API │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ DeepResearchAgent (Orchestrator) │
│ • Session management • Workflow execution │
│ • Error handling • Result aggregation │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ LangGraph Workflow Engine │
│ │
│ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ │
│ │ Init │──▶│Query │──▶│Search│──▶│Analyze│ │
│ └──────┘ └──────┘ └──────┘ └──┬───┘ │
│ │ │
│ ┌─────────────────────┘ │
│ │ [Routing Decision] │
│ ┌───────▼────────┬──────────────┐ │
│ │ │ │ │
│ ┌────▼───┐ ┌────▼────┐ ┌───▼────┐ │
│ │Continue│ │Connect │ │Finalize │ │
│ │(Loop) │ │(Graph) │ │Report │ │
│ └────────┘ └────┬────┘ └─────────┘ │
│ │ │
│ ┌────▼────┐ │
│ │Synthesize│ │
│ └─────────┘ │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ Service Layer │
│ ┌──────────────┬──────────────┬──────────────┐ │
│ │OpenAI Service│Claude Service│Search Service│ │
│ │• Web Search │• Reflection │• Parallel │ │
│ │• Entity Ops │• Query Gen │ Execution │ │
│ │• Graph Build │• Analysis │• Rate Limit │ │
│ │• Synthesis │ │ │ │
│ └──────────────┴──────────────┴──────────────┘ │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ Data & Observability Layer │
│ ┌──────────────┬──────────────┬────────────────┐ │
│ │State Models │Audit Logger │Operational Log │ │
│ │(Pydantic) │(JSONL) │(Structured) │ │
│ └──────────────┴──────────────┴────────────────┘ │
└─────────────────────────────────────────────────────────┘
- Investment Due Diligence: Comprehensive background checks on executives, board members, major shareholders
- Vendor Risk Assessment: Third-party due diligence, supply chain risk analysis
- Compliance & AML: PEP screening, sanctions checking, adverse media monitoring
- M&A Due Diligence: Target company investigation, key person analysis
- Background Verification: Executive hiring, partnership evaluation
- Investigative Research: Fraud investigation support, litigation intelligence
- Python 3.11+
- OpenAI API Key
- Anthropic API Key
git clone <repository-url>
cd deepagents
pip install -r requirements.txt
# Configure API keys
cp .env.example .env
# Edit .env and add your API keys# CLI execution
python -m src.main "Elizabeth Holmes" --context "Former CEO of Theranos" --max-depth 3
# Set custom search depth
python -m src.main "Bill Hwang" --max-depth 7Edit config/models.yaml to change models without code changes:
workflow:
max_search_depth: 5
max_queries_per_depth: 10
max_concurrent_searches: 5
query_generation:
provider: anthropic
model: claude-sonnet-4-5-20250929
temperature: 0.3
web_search:
provider: openai
model: GPT-4o# Required
OPENAI_API_KEY=your_key
ANTHROPIC_API_KEY=your_key
# Optional (for observability)
LANGFUSE_PUBLIC_KEY=your_key
LANGFUSE_SECRET_KEY=your_keyEach research session generates:
-
JSON Report (
reports/sess_*_report.json)- Executive summary & risk level
- 18 comprehensive sections
- Risk assessment with severity levels
- Entity relationship graph
- Source attribution and recommendations
-
Audit Log (
logs/sess_*.jsonl)- Complete immutable event trail
- All LLM calls with tokens/costs
- Compliance-ready format
-
Console Output
- Real-time progress updates
- Summary metrics and statistics
Typical session (depth=5, 10 queries/iteration):
- Duration: 6-8 minutes
- Queries: 40-50 total
- Cost: $2-5 USD
- Entities: 30-50 discovered
- Sources: 100-150 processed
Performance Modes:
- Fast (depth=2): 2-3 minutes
- Balanced (depth=5): 6-8 minutes
- Quality (depth=7): 12-15 minutes
| Document | Description |
|---|---|
| QUICK_START.md | Installation, usage, configuration, troubleshooting |
| SOLUTION_DESIGN.md | Architecture, workflow, technical decisions, implementation details |
Convert JSON reports to professionally formatted PDFs:
# Convert all reports to PDF (default)
python convert_to_pdf.py
# Convert latest report only
python convert_to_pdf.py --latest
# Convert specific session
python convert_to_pdf.py --session sess_20251208_212522Features:
- Professional formatting with color-coded risk levels
- Executive summary, findings, entity networks
- Comprehensive analysis sections
- PDFs saved alongside JSON files in
reports/
deepagents/
├── config/
│ └── models.yaml # Model configuration
├── src/
│ ├── main.py # Entry point & DeepResearchAgent
│ ├── agents/
│ │ ├── graph.py # LangGraph workflow definition
│ │ ├── nodes/ # 6 workflow nodes
│ │ └── edges/ # Routing logic
│ ├── services/
│ │ ├── llm/ # OpenAI & Claude services
│ │ └── search/ # Search execution
│ ├── models/ # Pydantic data models & state
│ ├── prompts/ # Modular prompt templates
│ ├── observability/ # Dual logging system
│ ├── config/ # Settings management
│ └── utils/ # Helper functions
├── docs/ # Documentation
├── tests/ # Test suite with evaluation personas
├── logs/ # Execution / Audit logs (JSONL)
└── reports/ # Generated reports (JSON)
The following subjects have been researched and documented:
| Subject | Session ID | Report Generated | Execution Logs |
|---|---|---|---|
| Andrew Ng | sess_20251209_095140 | ✅ JSON + PDF | ✅ JSONL |
| Bill Hwang | sess_20251208_212522 | ✅ JSON + PDF | ✅ JSONL |
| Isabel dos Santos | sess_20251208_211141 | ✅ JSON + PDF | ✅ JSONL |
| Adrian Cole | sess_20251208_205804 | ✅ JSON + PDF | ✅ JSONL |
| Dr. Lena Voronina | sess_20251208_202514 | ✅ JSON + PDF | ✅ JSONL |
| Elon Musk | sess_20251208_201533 | ✅ JSON + PDF | ✅ JSONL |
All research reports include comprehensive due diligence analysis with entity graphs, risk assessments, and evidence-based findings.