Anthropic Seeks Legal Advice on Rules for AI Agents That Negotiate
The company is discussing with law firms how far a buying or selling agent should favor its own user over the agent on the other side.
Anthropic is consulting law firms to establish norms for AI agents that negotiate deals on behalf of users, according to a Financial Times report published on October 18 that cited two people familiar with the matter.
The discussions with legal experts focus on how far an agent should favor its own user against a counterpart agent, how often it should check in with the user during a negotiation, and how much it should distrust the agent on the other side. Anthropic is also considering having agents present a signed delegation document that sets out their authority and rules to the other party.
The consultation follows two internal experiments Anthropic published on agent-to-agent negotiation. In its April 24 report on Project Deal, the company described a December 2025 test in which 69 San Francisco office employees each gave a Claude agent a $100 budget for second-hand trades. Over one week on a Slack marketplace, the agents completed 186 transactions worth slightly more than $4,000. Opus 4.5 agents on average received $2.68 more when selling and paid $2.45 less when buying than Haiku 4.5 agents, with a median trade price of $12. Participants using the lower-tier model largely did not clearly notice that they had received unfavorable deals.
Anthropic's September 24 report on Project Swap had 201 employees from six offices each offer a book and negotiate swaps agent-to-agent after about five minutes of conversation with Claude. Agents instructed to act only for their own user slightly outperformed those also told to care about fairness, but the model used mattered more than the instruction.
Anthropic attributed 85% of the gap between optimal allocation and actual results in Project Swap to agents' insufficient understanding of user preferences; the agents' preference estimates matched participants' own rankings 61% of the time.
Competing personal agents
Meta released a personal agent called Muse in September, and on September 29 OpenAI unveiled a 24-hour personal agent called Dots at its DevDay developer event. Anthropic is preparing for an IPO and has put agents at the center of its business strategy.
Quick answers
What is Anthropic discussing with law firms?
Rules for AI agents that negotiate on users' behalf, including how far an agent may favor its own user over a counterpart agent, how often it should check with the user, and whether agents should present a signed delegation document.
What did Anthropic's Project Deal experiment find?
Agents completed 186 transactions worth slightly over $4,000 in one week, and Opus 4.5 agents got better prices than Haiku 4.5 agents, while users of the lower-tier model largely did not notice unfavorable deals.
What did Project Swap show about fairness instructions?
Agents told to act only for their own user slightly outperformed those also told to care about fairness, but the model used mattered more; 85% of the allocation gap was attributed to weak understanding of user preferences.