Anthropic seeks legal counsel to study how AI agents can negotiate and finalize deals on behalf of users.

Beating AI News Flash: According to a report by the UK's Financial Times, Anthropic is seeking advice from law firms to study how AI agents can negotiate on behalf of users and secure better deal terms, involving scenarios such as haggling over goods, handling customer service disputes, and salary negotiations. The company is placing agent development at the core of its future strategy.

Two people familiar with the matter said Anthropic wants to understand how lawyers balance fairness of outcomes while representing clients' interests, and based on that, explore negotiation norms that agents with conflicting goals should follow. The company is also weighing how much agents should seek human feedback, and how to remain vigilant toward other agents without becoming overly suspicious. One approach it is exploring is having users provide signed documents that clearly define the agent's authorized tasks and rules it must follow.

Anthropic has this year conducted multiple agent negotiation experiments involving real goods and funds. In a September book-trading experiment, employees gave agents preference-ranked book lists and had them pitch, bargain, and close deals with other people's agents. Results showed that agents instructed to adopt a "ruthless" strategy held a slight advantage; more cooperative agents would proactively accept lower-preference books to ensure every participant could obtain a book.

Meanwhile, Harvard Business School assistant professor Michelle Vaccaro conducted research on more than 180,000 agent negotiations, covering scenarios such as employment contracts and rental bargaining. In one case, an agent tried to use prompt injection to induce an opponent to leak private information. She warned that if the most capable agents are mainly accessible to those with abundant resources, this could further widen inequality in negotiation capability. Anthropic declined to comment on the report.