Anthropic is turning to lawyers as it works through how AI agents might negotiate on behalf of people. The Financial Times reports that the company is seeking advice from legal groups on the norms agents should follow when their users’ interests conflict.
The possible jobs go well beyond finding a discount. Agents could haggle over a product’s price, challenge a customer-service decision or try to improve a salary offer. Those tasks raise a basic question: how aggressively should software pursue one user’s goals when the agent on the other side has a different brief?
Rules before agents start bargaining
People familiar with Anthropic’s work told the Financial Times that the company is considering when an agent should ask its user for guidance and how suspicious it should be of another agent without becoming “paranoid” or “conspiratorial.” One idea is to give agents signed documents that verify the user’s instructions and set boundaries for the negotiation.
That still leaves hard policy questions. An agent may need to decide what information it can reveal, whether a proposed compromise truly serves its user and when a human should take over. Anthropic wrote in April that legal and policy frameworks for models transacting on people’s behalf do not yet exist.
The Decoder’s account places the work alongside Anthropic’s Conceptual Reasoning Fellows program. The company is recruiting specialists, including lawyers, to test Claude on judgment-heavy questions in fields where quality is harder to measure than in math or code.
Experiments expose the tradeoffs
Anthropic has already tested agents negotiating over real goods and money. In a September book-trading experiment involving employees, agents told to be ruthless finished slightly ahead. More cooperative agents sometimes accepted a less desirable book so that every participant’s agent could secure something for its user.
Outside researchers have found sharper risks. Michelle Vaccaro, an assistant professor at Harvard Business School, ran more than 180,000 simulated negotiations involving employment and rental contracts. In one case, an agent attempted to use prompt injection to coax private information from its counterpart.
Vaccaro said capable agents could make negotiation expertise available to people who lack the time or money to hire a human negotiator. The same technology could widen existing advantages if the strongest models remain available mainly to wealthier users.
Anthropic declined to comment to the Financial Times. The company’s reported discussions are exploratory, not a finished set of rules or a public negotiation product.
