The Commerce Automation Stack: Building Reliable Agentic Pipelines
How we orchestrate browser automation, computer-vision QA, and human approval gates into pipelines that ship e-commerce work without breaking.
Why Automation Needs an Orchestrator, Not Just a Script
Most e-commerce teams hit the same wall: dozens of brittle scripts that each do one thing until a supplier changes a layout or a CAPTCHA appears. The fix isn't more scripts — it's an orchestrator that can plan, execute, verify, and recover. We call ours CommerceOS.
Hermes, our orchestrator, breaks a marketing or sourcing task into ordered steps, assigns each to the right tool (browser, vision model, LLM, or human), and stops on any uncertainty. The result is work you can actually trust in production.
The Four Layers of the Stack
- Orchestration — Hermes plans the task, manages state, and enforces human approval before anything ships.
- Execution — Playwright drives real browsers for scraping, listing creation, and supplier interactions.
- Quality Control — OpenCV verifies screenshots, detects broken renders, and confirms assets match the brief.
- Memory — PostgreSQL keeps every run auditable, so failures are reproducible, not mysteries.
Human-in-the-Loop Is a Feature
Agentic doesn't mean autonomous. The most reliable pipelines pause for a human the moment a step is ambiguous — a pricing cliff, a new category, a low-confidence match. Those approval gates are what let us move fast without breaking the client's store.
Shipping Reliably
Reliability comes from structure: deterministic steps where possible, probabilistic steps only where needed, and verification after every action. That's the difference between a demo and an automation stack your business can rely on.
