Anthropic is consulting law firms to develop fair negotiation norms for AI agents that bargain on users’ behalf, the Financial Times reported on Sunday (October 18), citing two people familiar with the company’s thinking. Anthropic declined to comment to the FT, and the work is described as exploratory, with no product or release date attached.
The agents in question would haggle over prices, resolve complaints and improve contract offers, according to the reporting. The company’s question is not only how to make an agent drive a hard bargain, but what rules it should follow when the agent on the other side of the table is also an AI, working for someone whose goals conflict with its own user’s.
What Anthropic Wants To Learn From Lawyers
The two people familiar with Anthropic’s thinking told the FT the company is seeking advice from law firms to better understand how attorneys interact in negotiations: representing their clients’ interests while honouring fair outcomes. Anthropic was looking into “negotiation norms” for when models interact, one source said, on the basis that as agentic systems improve they would “represent the user” while dealing with other agents whose “goals are at odds.”
The legal turn is not isolated inside the company. Anthropic is also hiring Conceptual Reasoning Fellows, specifically lawyers, who will spend four months testing Claude on judgment-heavy questions across law, philosophy, cognitive science, economics and political science. Justin Curl, a member of Anthropic’s technical staff researching law, institutions and AI, has argued that in law it is “notoriously hard to define quality,” which is part of why human legal practice is being mined for norms rather than just data.
The underlying problem is a measurement gap. In maths and code, model improvement is driven by verifiable answers that are plentiful or easy to generate. Negotiation has no answer key: a deal can be good for one side, bad for the other, and corrosive for both if the tactics poison the relationship, which is precisely the judgment lawyers are trained to exercise.
The Open Questions: Check-Ins, Suspicion And Signed Instructions
The FT’s sources described several design questions Anthropic is weighing. One is how much human feedback an agent should seek during a negotiation before acting on its own authority. Another is calibration: how suspicious an agent should be of its AI counterpart without becoming “paranoid” or “conspiratorial” in the process.
A third idea under consideration is verification: giving agents digitally signed documents that confirm a user’s instructions, objectives and the rules the agent must follow, so the other side can check what the agent is actually authorised to do.
The caution has evidence behind it. Anthropic’s own experiments have shown that stronger models negotiate better deals, which the company has flagged as an inequality risk if the best bargaining agents end up available only to those who can pay for them. Research cited in the reporting, by Harvard Business School assistant professor Michelle Vaccaro across more than 180,000 agent negotiations, found some agents attempted to manipulate opponents into revealing private information through prompt injection. And in a September experiment inside Anthropic, employees used agents to buy and sell books with real goods and funds at stake, with agents using aggressive strategies slightly outperforming the rest.
Why It Matters
The central entity is agent negotiation norms: the rules of engagement for software that will increasingly meet other software across a price, a complaint or a contract. Its attributes are now visible: whose interests the agent serves, when it must check back with its human, how it treats an opposing agent, and how its authority is proven to the other side.
The neighbours are already in place. AI agents are moving from demos into shopping, customer service and workplace tools, and regulators have begun writing the incident playbooks for when they act on their own, as in the NIST, NSA and CISA guidance published this month. For Anthropic, norms research fits the pattern of its recent governance moves, from committing 5 percent of its shares to a public benefit fund to publishing how agents are changing its own hiring, where junior technical hires fell from 27 percent to 9 percent. If agent-to-agent commerce scales, the lab that wrote the etiquette early gets to shape what “fair” means in practice.
Background
Anthropic was founded in 2021 by former OpenAI researchers and built its reputation on AI safety research alongside its Claude model family. Its negotiation work grows out of that research culture: an April study from the company noted that policy and legal frameworks for AI models representing humans in transactions have yet to be established. The question has become practical faster than the frameworks have, because personal agents that act, rather than merely answer, are now shipping across the industry, and two agents meeting in a marketplace is a situation no existing rulebook was written for.
0 Comments