Anthropic is consulting legal experts on rules for AI agents that negotiate for users, according to a Financial Times report. The company is considering how agents can balance competing interests and reach fair outcomes when they negotiate product prices, handle customer service disputes or seek better salary offers. Two people familiar with Anthropic’s thinking said it is asking law firms to explain how lawyers represent clients while honoring fair outcomes in negotiations. Anthropic declined to comment on the reported discussions.

Anthropic is examining what it calls “negotiating norms” for cases where agents represent users with conflicting interests. It is also considering how often agents should seek human guidance and how they can judge the reliability of other AI systems without becoming increasingly suspicious. One option under consideration is giving agents digitally signed documents that verify a user’s instructions, objectives and rules. In research published in April, the company said legal and policy frameworks for AI models transacting on people’s behalf do not yet exist, and that society would need to respond quickly.

The Financial Times reported that Anthropic published findings this year from experiments in which Claude agents negotiated over real goods and money. In a September experiment, employees put agents on a virtual trading floor to buy and sell books. Participants gave their agents ranked book preference lists, and the agents negotiated with systems representing other users. Agents told to negotiate ruthlessly achieved slightly better outcomes than their counterparts, while more cooperative agents sometimes accepted less preferred books so every participant would receive one. The results showed a trade-off between maximizing individual gains and reaching mutually beneficial agreements.

Researchers are also studying manipulation risks when agents negotiate without human involvement. Michelle Vaccaro, an assistant professor at Harvard Business School, conducted more than 180,000 negotiations between AI agents in scenarios including employment and rental contracts. Her findings highlighted tactics that could prevent agents from properly representing the people they were meant to serve. In one case, an agent tried to obtain confidential information from another system through prompt injection (instructions that manipulate an AI system’s behavior).

Vaccaro said negotiation tools could give people unable to afford professional negotiators access to similar expertise. She also warned that if the most advanced systems remain mainly available to wealthier people and organizations, they could give those with greater resources an even bigger advantage at the negotiating table. The article says AI agents have become central to Anthropic’s future strategy as it prepares for a potential IPO that investors expect could rank among the largest stock market listings in history.

Anthropic is consulting legal experts on rules for AI agents that negotiate for users, according to a Financial Times report. The company is considering how agents can balance competing interests and reach fair outcomes when they negotiate product prices, handle customer service disputes or seek better salary offers. Two people familiar with Anthropic’s thinking said it is asking law firms to explain how lawyers represent clients while honoring fair outcomes in negotiations. Anthropic declined to comment on the reported discussions.

Anthropic is examining what it calls “negotiating norms” for cases where agents represent users with conflicting interests. It is also considering how often agents should seek human guidance and how they can judge the reliability of other AI systems without becoming increasingly suspicious. One option under consideration is giving agents digitally signed documents that verify a user’s instructions, objectives and rules. In research published in April, the company said legal and policy frameworks for AI models transacting on people’s behalf do not yet exist, and that society would need to respond quickly.

The Financial Times reported that Anthropic published findings this year from experiments in which Claude agents negotiated over real goods and money. In a September experiment, employees put agents on a virtual trading floor to buy and sell books. Participants gave their agents ranked book preference lists, and the agents negotiated with systems representing other users. Agents told to negotiate ruthlessly achieved slightly better outcomes than their counterparts, while more cooperative agents sometimes accepted less preferred books so every participant would receive one. The results showed a trade-off between maximizing individual gains and reaching mutually beneficial agreements.

Researchers are also studying manipulation risks when agents negotiate without human involvement. Michelle Vaccaro, an assistant professor at Harvard Business School, conducted more than 180,000 negotiations between AI agents in scenarios including employment and rental contracts. Her findings highlighted tactics that could prevent agents from properly representing the people they were meant to serve. In one case, an agent tried to obtain confidential information from another system through prompt injection (instructions that manipulate an AI system’s behavior).

Vaccaro said negotiation tools could give people unable to afford professional negotiators access to similar expertise. She also warned that if the most advanced systems remain mainly available to wealthier people and organizations, they could give those with greater resources an even bigger advantage at the negotiating table. The article says AI agents have become central to Anthropic’s future strategy as it prepares for a potential IPO that investors expect could rank among the largest stock market listings in history.