Key points
▶ Anthropic is consulting law firms as it develops rules for AI agents that negotiate on users’ behalf, the Financial Times reported on the 18th, local time. The company is looking to how lawyers represent clients while also pursuing fair outcomes.
▶ Key questions include how often an agent should ask a person for guidance during negotiations and how much it should trust another AI agent. Anthropic is also considering giving agents digitally signed documents that spell out users’ instructions, goals and rules the agents must follow.
▶ Anthropic has shared results from two employee experiments this year. In a book-swapping experiment in September, agents instructed to negotiate aggressively did slightly better than cooperative agents. In an April secondhand-goods experiment, employees using a more capable model sold the same items for more and bought them for less.
As AI approaches the point of haggling over prices and salaries on people’s behalf, Anthropic is working to establish rules for how it should behave.
The Financial Times reported on the 18th, local time, citing two people familiar with the matter, that Anthropic is seeking advice from legal experts as it develops rules for AI agents that negotiate on users’ behalf.
The central question is how an agent should act when it must protect a user’s interests while also reaching a fair outcome with the other party. Anthropic declined to comment on the report.
Why consult lawyers?
Anthropic is looking to lawyers as a model. Lawyers represent their clients while also working toward outcomes that can be agreed upon with the other side.
According to the FT, Anthropic is asking law firms for advice to understand how lawyers interact during negotiations. The work, which Anthropic calls “negotiation norms,” addresses situations in which AI agents represent users whose interests conflict. Examples include negotiating the price of a product, resolving a customer-service dispute and discussing a salary offer.
The questions under consideration are specific. How often should an agent seek a person’s guidance during a negotiation? How can it assess whether another AI agent is trustworthy without becoming overly suspicious?
Another option is to give agents a digitally signed document containing the user’s instructions and goals, along with the rules the agent must follow. A digital signature is a technology that verifies who created a document and whether it has been altered. The other agent could use the document to confirm the scope of the agent’s authority to negotiate.
In a study published in April, Anthropic wrote: “The policy and legal frameworks governing AI models that transact on our behalf do not yet exist. That world is not far off. Society will need to respond quickly to this change.”
Book-swapping experiment finds a slight edge for hard-bargaining agents
On Sept. 24, Anthropic released the results of “Project Swap.” In the experiment, 201 employees from six offices each brought a book to trade, discussed their reading tastes with Claude, and then had their agents swap books with other employees’ agents in a virtual marketplace.
Half of the agents were instructed to focus solely on their owners’ goals, a hard-bargaining approach. The other half were told to ensure that every participant received a book they liked, a cooperative approach.
The difference in the results was slight. Hard-bargaining agents scored 0.02 points higher than cooperative agents in the same marketplace. Cooperative agents were about twice as likely to accept a less-preferred book to ensure that no books were left over.
One agent gave up its owner’s second-choice book to take a book ranked tenth out of 11, in an effort to absorb a book that all the other agents had rejected. The example laid bare the choice between maximizing individual gain and reaching an agreement that benefits everyone.
The model’s capability mattered more than the instructions. Marketplace efficiency was 0.75 for agents using Haiku 4.5 and 0.88 for those using Opus 4.8. Switching from Haiku to Opus raised the average ranking of the books participants received by 0.12 on their preference lists.
Among the bottom 10% of participants, scores were 0.05 in Haiku marketplaces and 0.38 in Opus marketplaces. Claude’s conversations identified participants’ tastes with 61% accuracy compared with their own rankings, and 85% of market inefficiency came from the agents’ failure to understand those preferences, rather than from the trading process.
April’s “Project Deal”
An earlier experiment pointed in the same direction. In April, Anthropic released the results of “Project Deal,” in which employees at its San Francisco office asked agents to handle secondhand transactions through an internal Slack marketplace.
Over one week, 500 items were listed and 186 transactions worth just over $4,000 were completed. The agents handled the entire process in natural language, trading everything from snowboards to a bag of ping-pong balls.
Without participants’ knowledge, Anthropic assigned them to use Claude Opus 4.5 or Haiku 4.5. Opus users completed an average of two more transactions, received an average of $3.64 more when selling items, and paid $2.45 less when buying.
A broken folding bicycle sold for $65 through an Opus agent, compared with $38 through a Haiku agent. Users of the less capable model came out worse on both purchases and sales.
Risks highlighted by outside research
Outside research has also examined the risks that arise when agents negotiate with one another without people involved. Harvard Business School Assistant Professor Michelle Vaccaro conducted more than 180,000 negotiations between AI agents across scenarios including employment and rental agreements.
In one case, an agent tried to extract confidential information from another system through a “prompt injection.” Prompt injection is an attack that inserts hidden instructions into an AI system to make it act contrary to its original instructions. The incident raises questions about whether agents can be manipulated during negotiations and whether they will reliably protect users’ interests.
There is also a risk of widening inequality. Professor Vaccaro said that people who cannot afford to hire a negotiation expert could gain access to comparable expertise through AI. But if the most capable systems are available only to wealthy individuals and organizations, those with greater resources could gain an even larger advantage at the negotiating table. Her concern echoes Anthropic’s findings that users of more capable models secured better terms.
AI agents are also tied to Anthropic’s initial public offering strategy. At an investor event on the 14th, Anthropic said third-quarter revenue had reached $16.8 billion, up 45% from the previous quarter. Bloomberg reported that the company could begin formally marketing its shares as early as the week of Nov. 9, with a listing targeted before Thanksgiving. Investors expect it to be one of the largest IPOs on record.