FOURWEEKMBA EXPLAINER · FROM THE NEWSROOM · UPDATED 15 NOVEMBER 2026
The short answer: Information design is the choice of what each AI agent is told about the others in a market: nothing, a rough signal such as a ranking, or the exact numbers. A Stanford-led study posted on 5 November 2026 tested that choice on ten frontier models trading, bargaining and bidding against copies of themselves. More detail was not reliably better. In double auctions, exact values lowered the market’s realised gains for nine of the ten models; Claude Opus 5.5 agents fell from 94.6% to 64.1% of the gains available. In two-agent trade, three models did better with exact values and seven did worse. In bargaining, seven did better. The authors’ conclusion: “model selection and information design must be co-optimized”.
Free to share with a link to this page. Data: Erdogan et al., arXiv:2611.07667. The figures are the gains realised by the whole market of agents, as a share of the maximum available, not one agent’s share.
When an agent buys or sells for someone, it has to know its principal’s limits: the most it can pay, the least it can accept. The paper’s point is that the agent on the other side should not necessarily see the same numbers:
“information an agent needs to represent its principal is not necessarily the same information that should be revealed to its counter-party.”
Economists have studied this for decades. The paper cites Blackwell (1953): a lone decision-maker cannot lose from more information. In a negotiation, though, information changes what the other side does too, so more of it can make the group worse off (Hirshleifer, 1978; Bergemann and Morris, 2016). What nobody knew is how today’s language models behave under each rule. The authors call their approach “empirical information design”: keep the models, markets and rules fixed, and change only what the agents are told.
The three disclosure levels
Least. Each agent knows only its own values or costs.
Coarse. A rough signal about the others: their ranking of the goods (trade), whether a deal is possible at all (bargaining), or the order of buyers by value and sellers by cost (auctions).
Full. The exact values and costs of the other side.
The authors, from Stanford University, the University of Maryland and All Purpose AI, ran this on ten models: GPT-5 mini, GPT-6 Luna, GPT-5.1, GPT-6 Sol, Gemini 3.7 Flash, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 5.5, Grok 4.3 and Grok 4.7. In each run, separate copies of the same model played every role and each was told to maximise its own gains. Because the researchers set the true values, they could score every outcome against the best one possible.
What happened in each market
Auctions: good trades priced out. Three buyers and three sellers each submit one price, once. With exact values, the auction still paired the right buyers with the right sellers: the pairs could have produced over 99% of the gains for Claude Opus 5.5 and Grok 4.7. The trades failed on price. “full disclosure leads sellers to ask for more above their costs in every model”, and buyers offered further below their values in every model except Gemini 3.7 Flash, the only model that did slightly better (+1.9 points).
Trade: it depends on the model. Two agents swap two goods over up to eight messages, across 90 markets per model and condition. Exact values helped three models and hurt seven.
The largest gain from exact values was Grok 4.3’s (+26.9 points), the largest loss GPT-5 mini’s (-10.2), per the paper’s Figure 1.
The losses came with grabbier offers. The paper counts “extractive” counteroffers, ones that raise the proposer’s gain while shrinking the total. In markets where one side had more room to extract, exact values raised GPT-6 Sol’s from 0.63 to 1.50 per negotiation and Claude Opus 5.5’s from 0.63 to 1.13.
Bargaining: more deals, less even splits. A seller and a buyer haggle over one good. Exact values often made agreement more likely but the split more lopsided. At four messages, GPT-5.1’s agreement rate rose from 70% to 93.3%, while its balance score among agreements fell from 0.84 to 0.24. In short exchanges, whoever moved first took more: with two messages and exact values, a GPT-5.1 seller who opened kept 97.3% of the gains, and one who responded kept 0.4%.
Our count from the paper’s Figure 1 (full minus coarse disclosure per model). Bargaining is measured at six messages.
When more detail helps
Harder search problems
With ten goods instead of two, exact values improved realised gains for nine of the ten models. GPT-5.1 rose from 63.8% to 80.4% of the maximum. When there are many possible deals, detail helps agents find the good ones.
A neutral mediator
When a mediator that maximises total gains received the exact values and proposed a deal both sides had to accept, full information gave near-optimal outcomes on the same markets where direct haggling did better with less.
More conversation
Going from two to twelve messages added 30.6 points for GPT-5.1 under the least information, against 2.8 points under full disclosure. Talking can partly replace information given up front.
Mixed teams behaved in a similar way. When different models faced each other, all six auction pairings did better with coarse information; full disclosure cut their scores by 4.9 to 14.6 points.
What this means for agent marketplaces
Agents already negotiate in tests. The paper opens with Anthropic’s Project Deal of December 2025, where about seventy people let Claude agents bargain for them and the agents closed 186 deals involving more than $4,000, as the paper describes it. As more of this moves to production, someone has to write the disclosure rule. As we read the study:
Treat disclosure as a design setting. The same models in the same markets moved by tens of points when only the information changed.
Test the model and the rule together. The best rule for one model was the worst for another: exact values lifted Claude Sonnet 5 in trade and cut GPT-6 Sol.
Watch the gap between found and closed deals. In auctions, agents found the right pairs and then priced them out. Measure both.
Coarse signals are often enough to match. Rankings let agents find good trades in simple markets without arming them to over-ask.
Detail can go to a referee instead. Giving exact values to a mediator, not to the bargainers, gave near-optimal results in the paper’s test.
What is not established
These are simulated markets with values set by the researchers, few agents and copies of one model on each side in the headline results. All main runs asked for low reasoning effort. The authors say real agents “may face uncertain preferences and changing goals” and that studies with deployed agents are needed. The paper is a preprint; we have not seen a peer-reviewed version. Code and data are public on GitHub, per the paper.
FAQ
What is information design in multi-agent AI?
The rule for what each agent learns about the others: nothing, a coarse signal such as a ranking, or exact values and costs. The Stanford-led paper shows the choice can change market outcomes by tens of percentage points.
Does giving AI agents more information improve outcomes?
Not reliably. Exact values helped seven of ten models in bargaining but only three in two-agent trade and one in double auctions, by our count from the paper’s Figure 1.
Why did full information hurt the auctions?
Sellers asked for more above their costs and buyers offered further below their values, so beneficial trades failed to clear, even though the right buyer-seller pairs were formed.
Which models were tested?
GPT-5 mini, GPT-6 Luna, GPT-5.1, GPT-6 Sol, Gemini 3.7 Flash, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 5.5, Grok 4.3 and Grok 4.7.
What should builders of agent marketplaces do?
The authors conclude that model selection and information policy should be chosen together. Test each model under each disclosure rule on your own objective before going live.
Source: Mete Erdogan, Mirac Suzgun, Aneesh Pappu, Guowei Xu, Batu El, Andrew A. Schoen, Furong Huang, John P. Dickerson, Mert Pilanci and James Zou, “Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies”, arXiv:2611.07667v1, 5 November 2026 (abstract, Sections 1-5, Figure 1, appendix). Read 15 November 2026. Quotes are verbatim from the paper. Model counts from Figure 1 are ours; the per-model values are the paper’s.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
