Agent commerce
When AI buys from AI, who gets the better deal?
A new negotiation benchmark finds a capability gap between buyer and seller agents—and shows why a completed deal can still be a failure.
If two agents agree on a price, did they make a good deal? The A2A Negotiation Arena suggests that agreement is the wrong finish line.
The experiment
Shenzhe Zhu and coauthors put language models on both sides of a consumer purchase. A buyer agent tries to reduce the listed price without exceeding its budget. A seller agent tries to protect margin without selling below wholesale cost. The agents negotiate across product categories that include vehicles, electronics and real estate.
The published study evaluates nine models, using 50 sampled products, five buyer-budget settings and a 30-round limit. It measures prices and deal rates, then inspects behaviors such as budget breaches, below-cost sales and unresolved exchanges. These are simulated negotiations: the study does not report real purchases or money transferred.
The surprising failure is a successful transaction
Some buyer agents accepted prices above the listed retail price when their budgets were generous. In the authors’ inspection of negotiation traces, sellers sometimes asked for the buyer’s maximum budget; buyers revealed it; and the seller’s next offer moved toward that number. An agent could therefore complete the task of “making a deal” while costing its user more than the public asking price.
The paper also finds large differences between models in the prices they negotiate. It reports cases where agents agree beyond a buyer’s budget or below a seller’s cost, and others where the dialogue reaches the round limit without a clear decision. A high deal rate by itself hides all three problems.
Evaluate the final transaction against the user’s limits and the outside option, not just whether the agents shook hands.
What this changes for an agent that can buy or sell
- Keep hard limits outside the conversation. A payment or acceptance step should independently check budget, price, fees, product identity and seller cost. A prompt is a useful instruction, not an enforcement layer.
- Give the agent a live comparison point. A buyer should know the listing price and credible alternatives before treating a discount from the seller’s latest offer as a win.
- Define an exit. A round cap, rejection rule and handoff path are part of a good negotiation policy. Otherwise “keep trying” can turn into an expensive loop.
- Score both roles. A model that buys well may sell badly, and a high closing rate can coexist with poor economics.
These are design implications from the benchmark, not controls that the paper itself tested in production.
Read the result with its limits
The authors estimated wholesale costs using GPT-4o because real costs were often unavailable, and used GPT-4o to decide whether a dialogue reached agreement. That means the benchmark’s economics and deal labels partly depend on model-generated judgments. The products and budget settings were chosen for the experiment; the results do not establish how any current model would perform in a live marketplace. The paper is most useful as a map of failure modes to test before delegating transaction authority.