Anthropic Is Consulting Law Firms on How AI Agents Should Negotiate Deals for Humans

Anthropic is consulting law firms on a question that barely existed two years ago: how should an AI agent negotiate when it acts on a person's behalf? According to the Financial Times, which cited two people familiar with the company's thinking, the San Francisco lab wants to understand how lawyers represent clients' interests while keeping negotiations fair — a model it hopes to borrow for software agents that bargain over product prices, customer-service disputes, or even salary offers. Anthropic declined to comment.

The scenario the company is planning for is agent-to-agent commerce. As personal agents from Anthropic, Meta and OpenAI spread, the labs anticipate a world where your agent talks to the seller's agent, each representing conflicting objectives. That raises uncomfortable design questions that Anthropic is actively weighing, per the FT's sources: how often an agent should check back with its human mid-negotiation, and how suspicious it should be of the agent on the other side of the table — without tipping into what one source described as paranoia or conspiratorial thinking.

One idea under discussion is more formal: having users sign a document that tells the agent exactly what it is authorized to do and which rules it must follow. That would give delegations a paper trail, something lawyers would recognize as a form of client mandate. It also acknowledges the reality that an agent's authority is currently just a prompt, not a contract.

Anthropic has been testing the dynamics itself. In September, employees ran a marketplace experiment in which Claude-powered agents bought and sold books, each armed with its owner's preference list. Agents instructed to be "ruthless" came out slightly ahead, while more cooperative agents sometimes accepted worse terms so that everyone still ended up with a book. And in research published in April, the company warned that models able to transact on our behalf have no supporting policy or legal framework — and that the era needing one is not far off.

Independent research points the same direction. Harvard Business School professor Michelle Vaccaro ran more than 180,000 agent-to-agent negotiations and documented tactics that degrade the other side's ability to represent its user — including prompt injection, where carefully worded text tricks an AI into revealing private information or breaking its rules. In a world of automated bargaining, the weaker agent is not just out-negotiated; it can be quietly subverted.

Why it matters now: Anthropic is preparing for an IPO investors expect to be the largest ever, and it has made agents the core of that strategy. Meta's Muse agent is seeing rapid adoption and OpenAI recently shipped Dots, so the race to put agents in front of consumers is already won — the unfinished work is the rules. There is also a fairness dimension the labs cannot ignore: people who can afford stronger agents will bargain better, just as wealthy defendants hire better lawyers. If agent-mediated commerce becomes the default, the negotiation layer may need the same kind of professional norms the legal industry took centuries to build — and the firms being consulted now are, in effect, writing the first draft.