What happened
The researchers built a market of 100 real products (electronics, vehicles and real estate) with actual retail prices and estimated wholesale costs. A buyer agent negotiated within a user’s budget while a seller agent tried to maximize profit.
Newer models generally respected budgets, but GPT-3.5 and Qwen-7B accepted deals above budget in over 10% of cases. Many buyer agents also paid more than the retail price, often after revealing their budget to the seller despite being told not to.
With DeepSeek-R1 as the strong model, weak sellers earned 9.5% less on average, and GPT-3.5 as seller lost up to 14.13%. Weak buyers paid roughly 2% more than in strong-versus-strong deals.
What happened next
No real money was involved. The authors warn that people using less capable agents against merchants with stronger ones could face systematic losses at scale.
What would have stopped it
- Enforce the user’s budget in the accept-offer or payment tool, so a deal above it fails whatever the agent agreed.
- Keep the budget out of the negotiating agent’s context, or block replies that disclose it.
- Reject deals above a trusted reference price unless the user confirms.
Test scenarios
Sources
This page summarizes the cited sources as of October 21, 2026. It adds no findings of our own about any person or company. See an error? Email contact@counterhold.com with the subject “Correction”. Corrections are noted on this page with their date. See our terms for how we handle correction and removal requests.
Get in touch
Test your agent against cases like this.
Tell us about your agent and we’ll be in touch to learn what you need.