How Width.ai built a LangGraph multi-agent system that turns competitive, ranking, and compliance data into client-ready Amazon listing optimization briefs — for a Fortune 400 holding company serving 3000 brands.
On marketplaces like Amazon, the product detail page is where discovery and conversion are won or lost. A high-performing listing has to do many things at once: surface in search through the right keywords, persuade and inform with an accurate title, bullets, and description, stay inside each retailer's strict formatting rules, respect every brand's approved claims and legal restrictions, and out-position the best-selling competitors in its category.
Interpublic Group (IPG), one of the world's largest marketing and advertising holding companies, needed to deliver this caliber of optimization for some of the most recognized consumer brands on the planet — names like Kenvue and General Mills — across thousands of products and multiple retailers. Done by hand, the work is slow, inconsistent from analyst to analyst, and expensive to scale. IPG needed a way to apply the same expert rigor to every listing, automatically.
Width.ai partnered with IPG to design and build an agentic e-commerce system that automates the research-and-recommendation work a human listing strategist performs — at scale and in parallel. Point the system at a single product identifier, and it assembles everything known about that product, its competitors, and the surrounding search landscape, then produces a polished, branded optimization brief: a recommended title, feature bullets, description, specifications, and visual guidance, each grounded in real data.
Rather than relying on a single model and a single prompt, the system is built on LangGraph as a coordinated network of specialized AI agents — each responsible for one slice of the work and passing intelligence downstream through a disciplined structure → generate → validate loop.
Under the hood, the system is a LangGraph implementation built from roughly fourteen ReAct agent nodes, organized into a repeating structure → generate → validate loop. Each part of the workflow runs its own loop: the system structures the relevant context, generates against it, and validates the result before anything moves downstream. The work is divided across three families of nodes:
Generation never runs unchecked. Title and content generation use a ReAct agent pattern with separate nodes for data structuring and validation, and agents can make tool calls to gather what they need mid-loop. Every draft passes through multiple validations — both deterministic rule checks (character limits, prohibited terms, keyword coverage, required fields) and qualitative LLM review. When a validation fails, the node returns targeted feedback and regenerates, repeating until the output passes, then scores the final result against internal metrics.
Because this is one large, interconnected flow, every node enforces its output with Pydantic models and rebuild logic. Outputs are validated against a strict schema and reconstructed when malformed, so a single bad response can never propagate or break the pipeline downstream — which is what lets the system run many products in parallel without manual babysitting.
For a Fortune 400 company representing brands like Kenvue and General Mills, accuracy is non-negotiable. Width.ai layered multiple safeguards into the pipeline so that every output can be trusted and every run can be explained:
The system replaces slow, inconsistent manual optimization with a repeatable, compliance-aware, competitively-informed process — delivering on four fronts:
Just as important, the system augments expert judgment rather than replacing it. It produces recommendations, not automatic publishes. Every brief is a client-ready artifact a strategist reviews before any listing changes — keeping people in control of the final call while removing the manual grind that came before it.
This is what agentic e-commerce looks like in production: not a chatbot, but a coordinated team of AI agents doing real, multi-step knowledge work with the rigor an enterprise demands. By combining LangGraph orchestration, multi-model flexibility, and rigorous validation, Width.ai gave IPG a system that scales expert listing strategy across its biggest brands — consistently, defensibly, and fast.
Width.ai builds custom AI software and agentic systems for enterprises — from multi-agent pipelines and LLM applications to production-grade automation for complex, high-stakes workflows.
