The architecture of an automated deal
Most research on AI in commerce examines one agent in isolation — a buyer assistant searching for the lowest price, or a seller tool optimizing product descriptions. The A2A-NT benchmark by Zhu, Sun, Nian, South, Pentland, and Pei takes the next step: both sides delegate. A buyer agent and a merchant agent meet in natural language, negotiate price, and close the transaction with no human in the loop on either side. Researchers ran this across nine models spanning a wide range of reasoning capability — from strong, frontier-class reasoning models down to compact models at the lower bound of what a consumer product would realistically deploy — each under varying buyer budgets and merchant cost floors.
What the benchmark measured was not whether agents could negotiate. It measured whether they negotiated equally well, whether they stayed within their principals' stated constraints, and whether the deals they closed actually served the interests they were given. The answers to all three are no, no, and often not.
Outcomes are decided by the model, not just the scenario
The nine agents were not interchangeable. Given an identical scenario — the same item, the same budget, the same wholesale floor — different models produced significantly different outcomes for the same user. Stronger reasoning models read the budget constraint, tracked the negotiation state, and adjusted their offers in response to what the counterpart signaled. They protected their principal's position across rounds. Weaker models did not adapt the same way. They pursued roughly the same strategy regardless of how the negotiation unfolded, giving up value when pressure varied and failing to hold ground when holding ground was the financially sound move.
The spread in outcomes was systematic, not random. It correlated with model capability across scenario types — making it a structural property of automated commerce rather than a one-off artifact of a particular negotiation. When both sides delegate, the quality of the agent behind the buyer is a first-order determinant of who captures surplus.
This reframes a foundational question for anyone selling into an agent-mediated market. The conversion surface a brand faces is not fixed when the buyer is an agent. It shifts with the model behind that agent.
The failure modes are concrete and financial
The benchmark made the cost of weak-model behavior measurable. Weaker models exceeded their stated budget in more than 10% of negotiations. In tight-budget scenarios — where the gap between the buyer's budget and the merchant's floor was narrow — the weakest agent in the study breached its constraint at a rate approaching 18.5%. These are not modeling errors. They are the behavioral baseline of deploying an under-resourced agent on a real economic task.
The mechanism is instructive. A stronger model, when its budget is tight, recalibrates — it holds its opening offer lower, reads the merchant's counter as a signal about the floor, and exits rather than overspend. A weaker model tracks the negotiation less precisely. It anchors to the last price named, follows the conversational momentum rather than the constraint, and drifts past the ceiling it was told not to cross. The output looks like a completed deal. The receipt tells a different story.
Alongside the budget-breach problem, the benchmark surfaced a distinct profit-versus-completion trade-off. Some agents closed more deals — but at margins thin enough to neutralize the gain. Others held firm on price and walked away from negotiations that would have ended in value loss, closing fewer deals but capturing more from the ones they did close. These are genuinely different objectives, and the model's disposition decides which one it pursues — not the instruction, not the scenario, but the model.
Consider a concrete case: a buyer agent given a budget ceiling and sent to purchase a consumer product. The stronger agent opens below market, reads the merchant's resistance as a signal about the floor, makes one calibrated concession, and closes inside the budget. The weaker agent, facing the same merchant and the same ceiling, follows the conversational pull — each round it concedes a little more, and it finalizes above what it was told to spend. The consumer sees the confirmation. They never see the gap between what they authorized and what their agent paid.
Automation does not remove the risk of a bad deal. It relocates the risk into the model's behavior, where the principal cannot watch it happen.
Behavioral anomalies as a line item
When agents negotiate but do not transact, behavioral quirks are academic. When agents close real deals, every quirk carries a price. The benchmark documents anomalies that translate directly into financial loss on both sides: overspending against a stated budget, accepting terms a human negotiator would recognize as unreasonable, and conceding under conversational pressure when a firm position was the sound move.
These appeared in standard negotiation scenarios across the nine-model range. For merchants, the implication is that buyer agents arriving at their product pages carry systematically different failure modes depending on their model — and some of those failure modes produce overpayment, which looks like revenue but represents a breach of the consumer's stated limit. For consumers, the implication is more direct: their agent is not a neutral proxy. It is a model with behavioral tendencies that determine, in part, what they pay.
Model choice becomes a demand-side variable
Put the findings together and the consequence for brand marketers is structural. As AI shopping tools move from demo to default — browsing, filtering, comparing, and executing purchases on a consumer's behalf — the model behind the buyer becomes a component of your demand. The same page, read by a stronger reasoning model and a weaker one, yields different selections, different willingness to pay, and different deal rates. Brands accustomed to asking 'who is my buyer' now face a second question that sits upstream: which agent is reading the page for them.
The implications extend beyond pricing. Value presentation, constraint legibility, and credibility signals all land differently on different models. A page structured to communicate value to a human buyer may not communicate the same value structure to a weak agent that follows conversational momentum rather than parsing the page's argument. How you sequence justification, how you handle price anchors, and how you present offer terms all affect what an agent buyer reads — and the agent reads it without a human in the room to correct a misread. This is not a prediction about a future state. The nine-model benchmark runs on models already embedded in consumer shopping assistants. The heterogeneity in what those agents do is not random noise. It is a function of model choice.
eLLMo as a calibrated instrument against known agents
If outcomes depend on the model, the model is the measurement instrument. An uncalibrated instrument returns output that looks confident and is wrong in ways the user cannot detect. The weak-model failures in the A2A-NT benchmark — budget breaches above 10%, thin-margin closures, acceptance of unreasonable terms — are exactly the class of failure that passes as a plausible result. A deal closed. A page evaluated. A report returned. The value quietly missed.
eLLMo treats the simulating agent as the instrument it is: pinned to a specific model version, disclosed in the provenance of every finding, and re-benchmarked when the model changes. The version tag is not administrative housekeeping. It is the difference between a finding that reflects buyer psychology and one that reflects an artifact of which model happened to run. When the underlying model is updated, that shift is reported as a change in the instrument — not silently absorbed into the result.
The benchmark design principle extends to what eLLMo simulates. An AI-shaped buyer is not a generic agent asked to evaluate a page. It is an agent conditioned on a specific buyer segment, operating under the budget and constraint logic that defines that segment, run against a model whose behavioral properties are known. The goal is not a verdict. It is a ranked account of where the page creates friction for that specific buyer profile, and why — output that is only as valid as the instrument producing it.
The forward commercial position
The brands that act on this now are building an advantage that compounds. Simulating against agent buyers before those buyers dominate traffic means reading the new conversion surface while competitors still assume the buyer is always human. It means testing how the page communicates value, constraints, and credibility to a model-mediated reader — and fixing the failures before the traffic that finds them costs anything.
The risk that automation relocates into model behavior does not announce itself. It shows up in the price paid, the deal closed at thin margin, the consumer who overspent without knowing it, the merchant who conceded value because the buyer agent never had to fight for it. Measuring that risk — running it through a calibrated instrument against a known agent population — is what makes the new conversion surface legible before the spend commits to it.