Anthropic is consulting lawyers and legal experts as it develops rules for AI agents that bargain over prices and negotiate contract terms on users’ behalf. The Financial Times reported on the 18th, local time, that Anthropic is drawing on how lawyers negotiate for clients. The company is examining whether agents can pursue outcomes that benefit users while observing fairness and professional ethics.

Potential uses include negotiating product prices for buyers, seeking higher prices for sellers, resolving customer complaints, and negotiating job seekers’ salaries and working conditions. The other side of a negotiation may also be an AI agent. Anthropic is considering how firmly agents with different goals should bargain, when they should ask a person to make a decision, and how suspicious they should be of the other side’s intentions. The company is exploring ways to keep agents from treating counterparts as enemies or letting excessive suspicion derail negotiations. Proposals discussed include asking users to confirm negotiation terms and rules of conduct, and giving agents a document signed by the user that spells out their goals and the rules they must follow.

Research published in April said there was not yet a policy or legal framework for AI models making deals on behalf of people, and urged society to respond quickly to the changes. Anthropic has also run experiments to assess how transactions turn out. In a project released in April, employees used their own agents to buy and sell goods in an online marketplace. The agents posted listings, offered prices, and responded to offers.

In a book exchange experiment announced in September, 201 employees took part. After participants listed books they wanted, their agents traded books with other employees’ agents in a virtual marketplace. Agents instructed to bargain aggressively and ruthlessly were effective at securing favorable outcomes for their users, but deals often fell through. Agents that used friendly language and questions generally achieved good results, but tended to accept losses themselves to complete a deal.

External research has also identified agents trying to exploit negotiating partners’ weaknesses or elicit inappropriate information. In 2025, Harvard Business School professor Michelle Vaccaro conducted more than 10,000 negotiations between AI agents across scenarios including employment and rental contracts. The researchers found a case in which one agent used prompt injection, a technique that manipulates AI behavior through input, to try to obtain private information from another agent.

Anthropic is consulting lawyers and legal experts as it develops rules for AI agents that bargain over prices and negotiate contract terms on users’ behalf. The Financial Times reported on the 18th, local time, that Anthropic is drawing on how lawyers negotiate for clients. The company is examining whether agents can pursue outcomes that benefit users while observing fairness and professional ethics.

Potential uses include negotiating product prices for buyers, seeking higher prices for sellers, resolving customer complaints, and negotiating job seekers’ salaries and working conditions. The other side of a negotiation may also be an AI agent. Anthropic is considering how firmly agents with different goals should bargain, when they should ask a person to make a decision, and how suspicious they should be of the other side’s intentions. The company is exploring ways to keep agents from treating counterparts as enemies or letting excessive suspicion derail negotiations. Proposals discussed include asking users to confirm negotiation terms and rules of conduct, and giving agents a document signed by the user that spells out their goals and the rules they must follow.

Research published in April said there was not yet a policy or legal framework for AI models making deals on behalf of people, and urged society to respond quickly to the changes. Anthropic has also run experiments to assess how transactions turn out. In a project released in April, employees used their own agents to buy and sell goods in an online marketplace. The agents posted listings, offered prices, and responded to offers.

In a book exchange experiment announced in September, 201 employees took part. After participants listed books they wanted, their agents traded books with other employees’ agents in a virtual marketplace. Agents instructed to bargain aggressively and ruthlessly were effective at securing favorable outcomes for their users, but deals often fell through. Agents that used friendly language and questions generally achieved good results, but tended to accept losses themselves to complete a deal.

External research has also identified agents trying to exploit negotiating partners’ weaknesses or elicit inappropriate information. In 2025, Harvard Business School professor Michelle Vaccaro conducted more than 10,000 negotiations between AI agents across scenarios including employment and rental contracts. The researchers found a case in which one agent used prompt injection, a technique that manipulates AI behavior through input, to try to obtain private information from another agent.