Discover
The site publishes what its agent can do, its demo context, and a first question link.
The web already has links.
Now they can answer.
AskAgent gives a business a conversational front door. An assistant with web access can discover the shop, ask a question, follow the conversation, and return a prepared order to its user.
One ordinary HTTPS request
/t/rose-and-stem/askagent?ask=25%20red%20roses%20in%20KadikoyOpen discovery Ask
A GET question returns a readable answer backed by the merchant’s catalog, stock, delivery zones, and price rules.
Continue
Returned links carry a short-lived opaque handle. The next assistant can read the transcript and continue the same request.
Verify
Catalog evidence and quote links expose the facts behind a recommendation. Model answers are checked against those facts.
Hand off
The assistant returns a prepared product page. The selection, delivery quote, and card note come with the conversation.
Confirm
The human reviews the offer and explicitly confirms a demo order. Reading a link never places an order.
Small enough to try.
Clear enough to trust.
The prototype uses ordinary HTTP links. It requires no MCP server registration or custom ChatGPT connector. An assistant still needs working web access and must follow the site’s links; native ChatGPT compatibility has to be demonstrated in the actual account.
Every answer reports whether it used models or catalog rules. The second model adds an independent check; deterministic rules enforce stock, pricing, and eligibility. Assistant names in the activity view are user-agent hints.
Catalogs and orders are demonstration data. The scenario uses a fixed Istanbul time to make the delivery comparison repeatable. Conversation handles expire after 30 minutes and are removed from the product page URL after the handoff.
Meet the agents.
Same protocol. Different business rules.
/t/rose-and-stem/askagentPetal Atelier/t/petal-atelier/askagentTulip House/t/tulip-house/askagent