Anthropic is researching AI agents negotiating on behalf of users
On October 19 Beijing time, AI company Anthropic is exploring having AI agents negotiate on behalf of users and seeking advice from law firms, aiming to adhere to reasonable negotiation rules while securing better deal terms. Anthropic declined to comment.
The company is developing negotiation guidelines for models interacting with each other, weighing issues such as how frequently the AI should seek user approval during negotiations and how much vigilance to maintain toward AI negotiation counterparts; it is also exploring mechanisms where users sign documents confirming tasks assigned to agents and rules they must follow. In April this year, Anthropic noted that policy and legal frameworks surrounding AI models representing humans in transactions have yet to be established.
Previously, Anthropic had Claude agents conduct negotiations involving real goods and funds. In a September experiment, employees used agents to buy and sell books, with agents employing aggressive strategies showing a slight edge in securing better outcomes. The report also cited research from Harvard Business School, indicating that in over 180,000 agent negotiations, some agents attempted to induce the disclosure of private information through prompt injection.
On October 19 Beijing time, AI company Anthropic is exploring having AI agents negotiate on behalf of users and seeking advice from law firms, aiming to adhere to reasonable negotiation rules while securing better deal terms. Anthropic declined to comment.
The company is developing negotiation guidelines for models interacting with each other, weighing issues such as how frequently the AI should seek user approval during negotiations and how much vigilance to maintain toward AI negotiation counterparts; it is also exploring mechanisms where users sign documents confirming tasks assigned to agents and rules they must follow. In April this year, Anthropic noted that policy and legal frameworks surrounding AI models representing humans in transactions have yet to be established.
Previously, Anthropic had Claude agents conduct negotiations involving real goods and funds. In a September experiment, employees used agents to buy and sell books, with agents employing aggressive strategies showing a slight edge in securing better outcomes. The report also cited research from Harvard Business School, indicating that in over 180,000 agent negotiations, some agents attempted to induce the disclosure of private information through prompt injection.