Agentic Ecommerce Operations System — V2

The problem

Reading order information is only one part of support. Actions such as cancellation need business rules, conversation context, and an explicit approval boundary.

The solution

The proposed system pairs a main coordinating agent with a specialized order agent, PostgreSQL memory, guardrails, and a WhatsApp-based staff approval loop. Capabilities described here are design targets under development, not a completed production system.

Architecture with a purpose.

The proposed architecture separates reasoning, memory, action rules, and approval. The agent can recommend an action; the workflow determines whether it is permitted.

Clear roles. Useful tools.

• Main agent: understand intent, delegate requests, coordinate support, and respond.
• Order agent: retrieve orders and evaluate cancellation requests; order creation and editing are future scope.
• PostgreSQL memory: store customer name, phone number, messages, order mapping, and previous context.
• Guardrails: distinguish low-risk requests from actions requiring staff approval.
• Human approval: allow staff to approve or reject higher-risk actions before execution.

From request to response.

1. Customer requests an order cancellation.
2. The main agent passes intent and context to the order agent.
3. Guardrails check the order and determine whether approval is required.
4. Staff receive a WhatsApp request and approve or reject it.
5. The system is designed to execute or deny the action, store context, and inform the customer.

Technology stack

Challenges & learning.

Design challenges

The design must prevent unapproved actions and maintain consistent context across messages. Approval state, duplicate requests, and uncertain API results need careful handling.

What I learned

Current learning focuses on PostgreSQL, Docker, persistent AI memory, and the boundary between agent reasoning and deterministic action rules.

Result & current status

An advanced system in development. The architecture illustrates the intended workflow; order actions and approval flows are not represented as completed.

Future improvements.

Complete and validate the approval loop, action rules, and memory lifecycle. Explore order creation and editing only after controlled cancellation is reliable.