Agentic AI Industry Use Cases — on AWS Bedrock AgentCore Harness

Industry agentic-AI applications rebuilt on AWS Bedrock AgentCore Harness — declarative,
fully-managed agents (no containers, no agent loop to write) with Gateway MCP tools,
a Bedrock Knowledge Base (S3 Vectors), managed Memory, and a single Cognito-secured
responsive PWA frontend.

finance-trading and healthcare-medical are deployed end-to-end and verified
(streaming chat, live tool calls, KB retrieval, cross-session memory, browser E2E, online
evaluations). The other 4 industries ship as deploy-ready templates on the same pattern.

Architecture

Browser (React PWA, Cognito JWT)
  │
  ├── chat ──► CloudFront /agent/* ──► AgentCore data plane: InvokeHarness (streaming)
  │                                       │ customJWTAuthorizer validates the Cognito token
  │                                       ▼
  │                                    Harness (declarative agent, Claude Sonnet)
  │                                       ├── AgentCore Gateway (MCP) ──► 5 Lambdas ──► DynamoDB
  │                                       │      market-data / portfolio / risk / trading / kb
  │                                       ├── kb tool ──► Bedrock Knowledge Base (S3 Vectors)
  │                                       ├── built-in browser + code interpreter
  │                                       └── AgentCore Memory (preferences / facts / summaries)
  │
  └── dashboards ──► API Gateway HTTP API (Cognito JWT authorizer) ──► Lambda ──► DynamoDB

Security posture: Cognito user pool (TOTP MFA, advanced security, 12-char passwords),
JWT verified at every entry point (harness customJWTAuthorizer + API Gateway authorizer),
WAF attached to CloudFront, least-privilege per-Lambda IAM, KMS on data stores, private
S3 behind CloudFront OAC, no long-lived secrets in code.

Repository layout

Path	What it is
`harnesses/<industry>/`	Declarative harness config template, memory strategies, system prompt
`tools/<industry>/`	Gateway Lambda tool handlers + MCP tool schemas
`tools/shared/toolkit/`	Shared dispatch, DynamoDB helpers, deterministic market simulator
`kb/<industry>/seed-docs/`	Knowledge-base seed documents (policies, product guides)
`skills/`	AgentCore Skills (git-sourced; wired post-merge)
`infra/cdk/`	6 CDK stacks: SharedSecurity, Auth, FinanceData, FinanceTools, Api, Web
`deploy/`	Orchestrator + idempotent scripts (gateway, memory, seed, render, smoke)
`web/`	Unified responsive PWA (Vite + React 19 + Tailwind + Amplify Auth)
`tests/`	Unit (pytest + moto), infra (CDK assertions), E2E (Playwright)

Deploy (flagship: finance)

Prereqs: Python 3.11+, Node 22+, AWS credentials, boto3 >= 1.43.51.

make setup            # venv + deps + CDK CLI
make test             # unit + infra tests
make deploy-finance   # CDK → seed → gateway → harness → memory → observability → smoke
make deploy-web       # build PWA → deploy WebStack → publish to CloudFront

deploy/deploy.py sequences everything and is idempotent — rerun it safely, or resume with
--from-step gateway. The AgentCore Harness Builder skill's scripts (preflight, validate,
create/update harness, invoke) are used underneath; set HARNESS_SKILL_DIR if the skill
lives elsewhere.

Verify

.venv/bin/python deploy/smoke_suite.py      # gateway tools, KB, memory (JWT end-user path)
cd tests/e2e && BASE_URL=https://<cloudfront> E2E_EMAIL=... E2E_PASSWORD=... npx playwright test

The six industries

Industry	Agent	Tools	Status
Finance Trading	`finance_trading_assistant`	16 domain + KB search	Deployed & verified
Healthcare Medical	`healthcare_medical_assistant`	16 domain + KB search	Deployed & verified (chat live)
Insurance Claims	template	claims / fraud / policy / settlement	Code-complete
Retail Inventory	template	inventory / forecast / supplier / pricing	Code-complete
Manufacturing Maintenance	template	equipment / prediction / maintenance / parts	Code-complete
Real Estate Valuation	template	valuation / market / investment / property	Code-complete

To deploy another industry: python deploy/deploy.py --industry <name> — the parameterized
IndustryStack plus the gateway/harness/memory scripts handle everything (healthcare was
brought online this way with zero new stack code).

Data honesty

Market data is a deterministic simulation (tools/shared/toolkit/market_sim.py) — stable
within a trading day, reproducible, no API keys. Every simulated payload carries
"source": "simulated" and the agent is instructed to disclose it. Orders are real writes to
the demo order book (DynamoDB) and genuinely mutate positions. Real-world lookups (news, SEC
filings) go through the harness's built-in browser tool.

Costs

Rough demo-scale monthly costs (us-east-1): CloudFront/S3/Lambda/DynamoDB on-demand ≈ $1–5,
S3 Vectors ≈ pennies, Cognito Plus per-MAU ≈ $0 at demo scale, no NAT/OpenSearch. The main
variable is Bedrock model usage during testing.

Disclaimer

Demo/reference architecture. Simulated financial data — not investment advice. Review
security, compliance, and cost for your own environment before production use.

License

Apache-2.0

agentic-ai-industry-use-cases