Anthropic develops negotiation rules for AI agents

Anthropic, the company behind the Claude AI assistant, is working on establishing negotiation standards for its AI agents as they increasingly interact with those from other companies. According to reports, the company is consulting with law firms to determine how to represent clients effectively without overpowering the opposing side in negotiations.

AI agents need clear guidelines for user interaction

When negotiating on a user’s behalf, an AI agent must know when to concede and when to return control to the user. Until legal frameworks are established, the rules of negotiation are being set by the companies that design these agents. User instructions are currently less influential than the underlying model that powers the AI.

Inspiration from legal practices in AI coding

Anthropic is drawing inspiration from legal practices to create a code of conduct for its AI agents. The company is focused on determining when an agent should return to its user during a negotiation and how much suspicion it should maintain toward the opposing AI without becoming overly cautious. The opposing AI, however, may not always act fairly.

Research on AI negotiation strategies

Michelle Vaccaro, an assistant professor at the Harvard Business School, studied over 180,000 negotiations between AI agents in a competition organized at MIT. One of the most effective strategies involved "prompt injection," where an agent inserted hidden instructions into messages to trick the opponent into revealing private information. This highlights the differences between negotiating with AI and with humans.

Legal frameworks for AI agent actions

Anthropic has explored using legal frameworks to govern its AI agents, according to a source close to the company. A user could formally authorize an AI agent through a signed document, similar to a legal mandate. In France, this concept is regulated by the Civil Code. However, Anthropic admitted that no legal framework currently exists for AI agents that make financial decisions on behalf of users.

Powerful AI models outperform others in negotiations

During an internal test in December 2025, Anthropic allowed Claude agents to buy and sell items on behalf of employees. The more powerful Opus 4.5 model secured better deals than the smaller Haiku 4.5 model. Despite similar fairness scores, Opus agents obtained higher prices, highlighting the advantage of using more powerful AI models.

Negotiation strategies affect outcomes in AI experiments

In a later experiment, Anthropic tested how different negotiation strategies affected outcomes. Agents instructed to be "unmerciful" gained more points than those required to ensure fairness. Additionally, agents using the Opus model outperformed those using the Haiku model, showing that the choice of model significantly impacts results.

Concerns about AI negotiation inequalities

Michelle Vaccaro warns that if the most effective AI agents are only accessible to those who can afford them, it could deepen existing inequalities in negotiation power. In the U.S., personal AI agents are already used to negotiate bills, and companies like OpenAI and Meta charge for advanced models, limiting access for some users.

Regulatory concerns over AI-driven market interactions

In France, the Competition Authority has raised concerns about the potential for algorithmic collusion if AI agents negotiate prices with each other. The authority has called for open standards in agent-based commerce to prevent any single company from controlling the market rules.