In B2B, much is decided in the quote. Three Lab projects work together: an agent answers price requests within the merchant's limits, a control layer keeps seller agents to the rules, and a benchmark measures what a negotiation strategy actually delivers.
| Context | Shopware Agentic Commerce Lab |
|---|---|
| Status | Three Lab projects |
| merchant-quote-agent-plugin | github.com/agentic-commerce-lab/merchant-quote-agent-plugin Licence MIT · Version 0.1.1 of 22 October 2026 |
| sales-agent-harness | github.com/agentic-commerce-lab/sales-agent-harness Licence MIT |
| rfq-bench | github.com/agentic-commerce-lab/rfq-bench Licence MIT |
| Checked on | 18 November 2026 |
Merchant Quote Agent
The extension for Shopware 6.7 answers price requests on B2B quotes. The merchant sets the maximum discount. Ordinary code checks that limit before and after saving, so the model cannot exceed it. Requests that are not about price go to a person. In draft mode, the agent only prepares replies. The project is in alpha and requires quote management from Shopware Commercial and your own model access.
composer require shopware/merchant-quote-agent-plugin
bin/console plugin:install --activate MerchantQuoteAgentPlugin
Sales Agent Harness
The harness is a control layer between a seller agent and the commerce systems. It is owned by the merchant and not tied to any agent framework. The agent searches products, creates cart drafts, summarizes them and hands over to checkout. The model does not decide prices, discounts, availability or commitments.
docker compose up --buildrfq-bench
rfq-bench is a reproducible benchmark for agents negotiating B2B requests for quotation. It measures the value an agent secures above its walk-away point. Every run has fixed, recorded conditions and immutable traces. It compares seller strategies such as Boulware or anchoring high, models in head-to-head duels and buyers with different priorities.
uv sync --extra dev
uv run rfq-bench run --overwrite
uv run rfq-bench report