Anthropic is asking law firms for lessons before giving negotiating agents their rules. The Financial Times reported on October 18, 2026 that, according to two people familiar with Anthropic's thinking, the company is consulting law firms on how lawyers represent clients in negotiations while ensuring fair outcomes, as it develops AI agents that would haggle over prices, resolve complaints and disputes, and secure better contract and salary offers for users, and interact directly with other AI agents. The report did not name the firms, and Anthropic has not published any written norms.
What needs defining is not wording, but the limits of authority
The hard items on the discussion list reported by the FT are not about phrasing. First, at which points in a negotiation must an agent stop and seek human feedback, and how much. Second, how suspicious an agent should be of a counterpart agent with conflicting goals, without sliding into paranoia or conspiratorial responses. Third, an idea the company has explored: giving agents a signed user document that verifies the instructions they received and the rules they must follow, the equivalent of a power of attorney in a human lawyer's hands. In the language of the legal profession this is familiar ground: how far an agent may go, when it must check back, and what proves the scope of its authority are the three basics of any agency relationship. In research posted in April, Anthropic said policy and legal frameworks for models transacting on users' behalf do not yet exist, that such a world is not far away, and that society would need to respond quickly; consulting law firms is part of that catch-up work.
Why it cannot wait: stronger models are already winning
Anthropic's own experiments add urgency. In September, employees ran an experiment in which agents bought and sold books on a virtual trading floor from preference lists; agents instructed to be ruthless achieved slightly better outcomes, while more cooperative agents knowingly accepted less-desired books so that every agent could return a book to its user. Harvard Business School assistant professor Michelle Vaccaro has run more than 180,000 negotiations between AI agents across scenarios including employment and rental contracts, and found novel tactics that undermine another agent's ability to represent its user, in one case using prompt injection to coax private information out of the other side. The FT also notes that stronger models negotiate better deals, which means people who can afford stronger agents would systematically come out ahead in bargaining. If the rules are not set before the products ship, the unfairness will not stay inside experiments.
At the same time, everyone is drawing lines around their agents: major US banks have already tightened agent account permissions, closing off actions such as delegated logins and transfers for now. Anthropic is making agents central to its future strategy ahead of a planned initial public offering, and the FT reports that adoption of Meta's personal agent product has risen sharply while OpenAI recently released a competing product called Dots. Negotiation norms sound like an ethics topic, but in product terms they are a trust question: whether users will hand bargaining, complaints and salary talks, all of which involve money, to a piece of software depends on that software first proving it knows where its limits are.
