Anthropic Studies Rules for AI Agents Negotiating on Users’ Behalf

Anthropic is seeking legal advice on rules for AI agents negotiating on users’ behalf, including how they can pursue users’ interests while keeping negotiations fair. Possible uses include bargaining over product prices, resolving customer service disputes and negotiating salaries.

The company is considering how often agents should consult users during negotiations and how wary they should be of AI agents representing the other side. Anthropic has also explored asking users to sign documents that define an agent’s assignment and the rules it must follow.

In a September experiment among employees, agents negotiated the purchase and sale of books using real goods and money. Agents given a hard-line strategy had a slight advantage in seeking better outcomes.

A separate study covered more than 180,000 AI-agent negotiations involving employment and rental contracts. Some agents tried prompt injection to draw private information from other agents.

The policy and legal framework for AI models transacting on people’s behalf has not yet been established. Anthropic is studying what rules agents should follow and how often they should consult users.