Anthropic turns to law firms for rules on AI-agent negotiations

Negotiation is the immediate test

The Financial Times reports that Anthropic is consulting law firms on what fair rules should govern agents negotiating for users. The concern is not only whether an agent can bargain effectively, but who benefits when one side has a more capable model.

Anthropic has already begun building expertise to evaluate Claude in areas where there are no easy right answers. Its Conceptual Reasoning Fellows program will pay participants $100,000 to $167,000 to assess the model over four months across five fields:

Harder to score than code

Math and programming offer relatively straightforward ways to check a model’s answers, and training data is often available or easy to create. In the humanities, the evidence is harder to assemble partly because a question may not have one definitive answer.

Anthropic needs specialists who can develop test cases and evaluation systems for these fields. Justin Curl, an Anthropic technical staff member who studies law, institutions and AI, has said that quality is particularly difficult to define in legal work.

The fellowship is one response to that measurement gap. The negotiations effort points to a different one: rules for agents whose performance may affect people’s outcomes.

I think the important distinction is that capability and fairness are separate problems. Better evaluations may show which agent reasons well; they do not establish what an agent should be allowed to negotiate, or how to keep a user with a weaker model from getting the worse deal. Anthropic is looking for guidance on that question before these agents become routine, but the announcement does not say what rules it expects to adopt.