{"text":[[{"start":7.95,"text":"Anthropic is in talks with legal groups about how to solve a novel problem: how AI agents should negotiate to secure the best deal for their human masters."}],[{"start":17.4,"text":"The San Francisco-based company is among the leading AI groups developing so-called autonomous agents, which are directed by users but then take independent actions to complete tasks."}],[{"start":28.95,"text":"The rise of personal agents is leading AI companies to contemplate a future where agents communicate with one another — and then negotiate on our behalf. This includes working out how such machines should haggle over the price of a product, resolve customer service disputes or even secure a better salary offer."}],[{"start":46.95,"text":"Two people familiar with Anthropic’s thinking said the company is seeking advice from law firms to better understand how lawyers interact in negotiations, representing their clients while honouring fair outcomes."}],[{"start":59.1,"text":"The move comes as Anthropic puts the development of AI agents at the heart of its future strategy, ahead of a planned initial public offering that investors believe could be the biggest stock market listing in history."}],[{"start":71.2,"text":"Setting the rules of engagement for agents has become increasingly more important as the technology is adopted more widely and becomes a key way that people and businesses use AI. "}],[{"start":81.9,"text":"Meta has seen a large jump in uptake of its personal agent tool, Muse, while OpenAI recently released a competitor product called Dots."}],[{"start":91.35000000000001,"text":"Anthropic was exploring “negotiation norms” when models interacted, according to one of the people familiar with Anthropic’s work, who added that as agentic systems improved, they would “represent the user” while encountering other agents whose “goals are at odds”."}],[{"start":108.05000000000001,"text":"The company was also balancing how much agents should seek human feedback in a negotiation, and how suspicious agents should be of AI counterparts without getting “paranoid” or “conspiratorial”, this person said."}],[{"start":120.15,"text":"Another idea Anthropic had explored was giving agents signed documents from the user verifying its brief and any rules to follow, one person close to the company said."}],[{"start":131.05,"text":"“The policy and legal frameworks around AI models that transact on our behalf simply don’t exist yet,” the company said in research posted in April. “Such a world isn’t far away. Society will need to move quickly to reckon with these changes.”"}],[{"start":144.4,"text":"Anthropic has also published findings from multiple experiments this year in which its Claude agents negotiated over real goods and money."}],[{"start":153,"text":"In September, Anthropic employees engaged in an experiment to buy and sell books, using AI agents on a virtual trading floor to “pitch, haggle and strike deals with other people’s agents”. The employees gave their agents a list of books ranked by preference."}],[{"start":168.75,"text":"The study found that agents that were instructed to be “ruthless” came out “slightly ahead” of other agents in securing a better outcome. Other more co-operative agents “made knowing sacrifices”, including accepting less-desired books to ensure each agent in the study ended up with a book to take back to the user."}],[{"start":187.75,"text":"Academics have also sought to tackle the problem. Michelle Vaccaro, an assistant professor at Harvard Business School, ran more than 180,000 negotiations between AI agents across scenarios including employment contracts and rental contract bargaining. "}],[{"start":201.65,"text":"It identified novel tactics that undermine another agent’s ability to represent the user. For example, in one case, an agent tried to coax a counterpart into revealing private information through a technique called prompt injection."}],[{"start":216.70000000000002,"text":"However, there are concerns that those who can afford more advanced AI models will gain the upper hand in any negotiation between agents. "}],[{"start":226.10000000000002,"text":"“Highly capable models could make negotiation expertise accessible to people who previously lacked the time or resources to develop those skills themselves or hire someone to negotiate on their behalf,” said Vaccaro. "}],[{"start":239.45000000000002,"text":"“On the other hand, if the strongest agents are available primarily to those with the most resources, they could reinforce existing advantages and widen inequalities in bargaining power.”"}],[{"start":251.25000000000003,"text":"Anthropic declined to comment."}],[{"start":260.6,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1792299810_9897.mp3"}