AgentCore Multi-Agent Financial Risk Assessment

A demo showing Amazon Bedrock AgentCore capabilities through a financial risk assessment use case:

• AgentCore Runtime -- 4 specialist Strands agents + 1 supervisor, each in its own runtime context
• AgentCore Memory -- Conversation persistence across assessment sessions via AgentCore Memory SDK
• MCP Tool Use -- Real FastMCP servers (credit bureau, employment, market data, compliance screening) consumed via MCPClient

Architecture

flowchart TD
  UI["React UI (dark theme)"] -->|SSE stream| API["FastAPI Backend"]
  API --> Supervisor["Supervisor Agent\n(Strands / Claude Sonnet 4)"]

  subgraph runtime [AgentCore Runtime]
    Supervisor -->|"agents as tools"| Credit["Credit Analyst\n(Claude Haiku 4.5)"]
    Supervisor -->|"agents as tools"| Income["Income Verifier\n(Claude Haiku 4.5)"]
    Supervisor -->|"agents as tools"| Market["Market Analyst\n(Claude Haiku 4.5)"]
    Supervisor -->|"agents as tools"| Compliance["Compliance Officer\n(Claude Haiku 4.5)"]
  end

  subgraph mcpServers [MCP Servers - FastMCP via stdio]
    Credit -->|MCPClient| CreditMCP["credit_bureau_mcp\n3 tools"]
    Income -->|MCPClient| EmployMCP["employment_mcp\n3 tools"]
    Market -->|MCPClient| MarketMCP["market_data_mcp\n2 tools"]
    Compliance -->|MCPClient| CompMCP["compliance_mcp\n3 tools"]
  end

  subgraph memoryLayer [AgentCore Memory]
    Supervisor --> Memory["AgentCoreMemorySessionManager"]
    Memory --> MemoryAPI["Bedrock AgentCore Memory API"]
  end

  CreditMCP --> Data[(mock_data.py)]
  EmployMCP --> Data
  MarketMCP --> Data
  CompMCP --> Data

Loading

Data flow (single assessment)

sequenceDiagram
  participant U as User / UI
  participant S as Supervisor
  participant CA as Credit Analyst
  participant MCP as credit_bureau_mcp
  participant M as AgentCore Memory

  U->>S: "Assess Sarah Chen for a $750K mortgage"
  S->>M: Retrieve session context
  M-->>S: Prior context (or empty)

  S->>CA: Delegate credit analysis
  CA->>MCP: get_credit_score(C-1001)
  MCP-->>CA: {score: 742, range: Good}
  CA->>MCP: get_credit_history(C-1001)
  MCP-->>CA: {history_years: 12, ...}
  CA->>MCP: get_debt_summary(C-1001)
  MCP-->>CA: {dti: 5.5%, ...}
  CA-->>S: {score: 82, rating: B+, ...}

  Note over S: Repeat for Income, Market, Compliance

  S->>S: Synthesize final assessment
  S->>M: Store session context
  S-->>U: Risk Score 82 — APPROVE

Loading

Quick start (local -- mock mode)

No AWS credentials needed. Uses pre-built responses for instant demo:

make install

# Terminal 1
make mock

# Terminal 2
cd frontend && npm run dev

Open http://localhost:5173

Real mode (Strands + Bedrock + MCP + Memory)

Requires AWS credentials with Bedrock model access:

# 1. Configure AWS
export AWS_REGION=us-east-1
aws configure  # or use SSO, profiles, etc.

# 2. (Optional) Create AgentCore Memory resource
make setup-memory
export AGENTCORE_MEMORY_ID=<id from output>

# 3. Run with real agents
make real

# 4. Frontend (separate terminal)
cd frontend && npm run dev

Deploy to AWS (CDK)

For deploying infrastructure to a staff account:

# 1. Bootstrap CDK (first time only)
cd infra && npx cdk bootstrap

# 2. Deploy IAM roles, ECR repos, SSM params
make infra-deploy

# 3. Build and push agent container
cd backend
docker build -t agentcore-risk-backend .
# Tag and push to ECR (URI from SSM /agentcore-risk/supervisor/ecr-uri)

Project structure

├── backend/
│   ├── server.py              # FastAPI + SSE streaming
│   ├── orchestrator.py        # Multi-agent orchestration (mock + real paths)
│   ├── mock_data.py           # Realistic financial data
│   ├── agents/                # Strands agent definitions
│   │   ├── credit_analyst.py  #   → credit_bureau_mcp
│   │   ├── income_verifier.py #   → employment_verification_mcp
│   │   ├── market_analyst.py  #   → market_data_mcp
│   │   └── compliance_officer.py  → compliance_screening_mcp
│   ├── mcp_servers/           # Real FastMCP servers (stdio transport)
│   │   ├── credit_bureau.py
│   │   ├── employment.py
│   │   ├── market_data.py
│   │   └── compliance.py
│   ├── requirements.txt
│   └── Dockerfile
├── frontend/                  # React + Tailwind dark-theme UI
│   └── src/
│       ├── App.tsx
│       └── components/
│           ├── ChatPanel.tsx
│           ├── AgentPanel.tsx
│           └── RiskDashboard.tsx
├── infra/                     # CDK infrastructure
│   ├── lib/agentcore-risk-stack.ts
│   └── bin/app.ts
├── scripts/
│   └── setup-agentcore-memory.py
├── docs/
│   └── DEMO-SCRIPT.md         # Talk script and demo walkthrough
├── Makefile
└── README.md

Environment variables

Variable	Default	Description
MOCK_MODE	true	true = simulated agents, false = real Strands + Bedrock
MODEL_ID	us.anthropic.claude-sonnet-4-20250514-v1:0	Supervisor model
HAIKU_MODEL_ID	us.anthropic.claude-haiku-4-5-20250501-v1:0	Sub-agent model
AWS_REGION	us-east-1	AWS region for Bedrock + Memory
AGENTCORE_MEMORY_ID	(empty)	AgentCore Memory resource ID (from setup script)

What's real

Component	Mock mode	Real mode
Agents	Simulated delays + pre-built responses	Strands SDK agents calling Bedrock
MCP Tools	Dict lookups returning mock data	Real FastMCP servers over stdio
Memory	Python dict (MEMORY_STORE)	AgentCore Memory API + local fallback
LLM calls	None	Claude Haiku 4.5 (sub-agents) + Sonnet 4 (supervisor)