AI Agents
Sales Agent
The full commerce conversation — browse, discover, order, reorder, cancel — as one continuous flow.
Manual- LangGraph.js
- GPT-5-mini
- Shopify GraphQL
- pgvector
- PostgreSQL
- Tools
- 8
- Verticals
- 5
Read the steps as a list
- trigger Delegated by orchestrator
- ai Sales agent loop · GPT-5-mini
- integration search_products · pgvector
Semantic search for a described need — “something for a gift”.
- integration filter_listings
Structured attribute filtering for non-retail catalogues — 3-bed in Maadi under 3M EGP.
- integration check_inventory · Shopify
Live stock. The only source of truth; never read from the vector index.
- condition In stock?
- action create_order
Claims store inventory first, then writes locally. Compensates on failure.
- action Offer alternatives
- output Order confirmed
The problem
Customers do not separate browsing from buying the way a system diagram does. Handing off mid-sentence between a “browse” agent and an “order” agent lost context and dropped sales.
How it works
Handles everything from “show me what you have” to a placed order. Browsing and ordering were originally two agents; splitting them turned out to be the single biggest routing failure point, so they were merged back into one.
The outcome
One agent, eight tools, and live stock as the only source of truth — never trusted from embeddings. Stock is claimed in the store before anything is written locally, with automatic compensation if that fails.