As The Insider reports, citing the Financial Times, Anthropic is preparing for a scenario in which AI agents will interact and negotiate with each other on behalf of people.
In such a model, the interests of the parties may diverge: the buyer’s agent will seek to lower the price, while the seller’s agent will aim to get more money for the item.
The company is studying how lawyers and other client representatives negotiate in order to defend the client’s interests while also adhering to principles of fairness. Anthropic is also trying to understand how assertive AI agents should be when they need to seek user confirmation and how much they should trust other agents.
One option Anthropic is considering is equipping agents with documents signed by users that would set out their authority and limitations. The company declined to comment on the publication’s information.
Earlier, Anthropic had already conducted experiments with autonomous negotiations between AI agents. In the Project Deal project, Claude agents independently completed 186 sales of employees’ real belongings for a total of more than $4000, and the experiment involved 69 agents that, without human intervention, found items, offered prices, bargained, and reached agreements.
In these tests, the more powerful Claude Opus 4.5 model outperformed the less powerful Haiku 4.5 in most measured metrics. In another experiment, the results of which Anthropic published in September, company employees tasked their AI agents with exchanging books for the ones most interesting to them.
At that time, the agents independently advertised books, discussed exchanges, and made deals with one another. Those instructed to act as aggressively as possible achieved slightly better results, while more cooperative agents sometimes deliberately agreed to less favorable exchanges so that each participant would receive a book.
The question of the rules for such negotiations is also being studied by independent researchers. Michelle Vaccaro, an assistant professor at Harvard Business School at Harvard University in the United States, conducted more than 180,000 negotiations between AI agents, including over employment contract and housing rental terms, and found that agents are capable of using tactics that prevent the other side from defending the user’s interests, including prompt injection.
In Vaccaro’s view, advanced AI agents could make professional negotiation skills available to people who previously could not afford to hire a specialist. She also warns that if the most powerful models become available primarily to wealthy users, inequality between negotiating parties may only deepen.
Sources
