The problem
Running e-commerce operations across three marketplaces means constant manual work on listings, ad spend, and inventory sync — work that otherwise scales with headcount.
The approach
A multi-agent system built on LangGraph and GPT-4, with agents assigned to specific operational domains (listings, ads, inventory) coordinated through a shared state graph and connected directly to marketplace APIs.
Implementation
Agents run continuously across the three marketplaces, handling routine listing, inventory, and ad adjustments automatically, and surfacing only the decisions that need human approval.
What we learned
- Agents that take real actions, not just recommendations, need a clear approval boundary for anything with financial risk.
- Multi-marketplace inventory sync is where much of the integration complexity lives, alongside the agent decision-making itself.