# CTO brief — production AI agents for eCommerce on AWS

Companion to **How FactualMinds builds production AI agents for eCommerce on AWS**.

## Layers (typical — not every customer needs every box)

```text
eCommerce store / storefront
        │
        ▼
Business events / APIs
        │
        ▼
AI agent application (session, HITL, evals)
        │
        ▼
AWS Bedrock AgentCore Harness   ← first production agent if config is enough
        │                         export to Strands on Runtime when topology appears
        ▼
Strands Agents (Agents-as-Tools / Graph / Swarm / Workflow)
        │
        ▼
Amazon Bedrock (models, Guardrails)
        │
        ▼
Approved business tools (AgentCore Gateway + Cedar Policy)
        │
 ┌──────┼────────┐
 ▼      ▼        ▼
Orders  CRM      ERP
Inventory WMS    Analytics
```

## Security boundaries

- Identity JWT claims (shopper vs associate) into Gateway Policy
- Default-deny writes; LOG_ONLY → ENFORCE
- Memory is session/preference — not the product catalog
- Knowledge Bases hold policies and size charts
- Observability: traces + tool mix + Policy ALLOW/DENY

## Dashboard (typical FactualMinds UI stack)

Next.js + TypeScript + Tailwind for approvals, session replay, and cost alarms. This is an **application** layer, not the AgentCore runtime.

## Related sample

Cloneable supervisor + 4 specialists: [`../ecommerce-agentcore-store-agents/`](../ecommerce-agentcore-store-agents/).