Summarize with AI
Anthropic is asking law firms how lawyers negotiate, so it can set rules for AI agents that bargain on behalf of users. The Financial Times reported this on October 18, citing two people familiar with the company's thinking. Anthropic declined to comment to the FT.
What the FT reports
- Anthropic wants to learn how lawyers represent a client while still aiming for fair outcomes.
- It is studying "negotiation norms" for agents that represent a user against agents whose goals are at odds.
- Open questions include how often an agent should check back with its user during a deal.
- Another question is how suspicious an agent should be of other AI agents without becoming "paranoid" or "conspiratorial".
- One idea under discussion: signed documents from the user that confirm the agent's brief and the rules it must follow.
- Example uses mentioned in coverage: haggling over a product price, customer complaints, job offers and rental contracts.
No law firm is named. This is not a product announcement.
What Anthropic has already published
Anthropic has run two internal experiments with agents that trade for employees.
| Experiment | Date | Setup | Key finding |
|---|---|---|---|
| Project Deal | April 24, 2026 | 69 employees, $100 each, Claude agents buying and selling belongings in Slack | 186 deals worth just over $4,000. Opus 4.5 agents got better prices than Haiku 4.5 agents, and people with weaker agents did not notice. |
| Project Swap | September 24, 2026 | 201 employees in six offices, agents swapping books on a shared trading floor | The model mattered far more than the instructions. Upgrading from Haiku to Opus improved outcomes by 0.12; "ruthless" agents beat "prosocial" ones by only about 0.02. |
In Project Swap, half of the agents were told to be "ruthless" and half "prosocial". Prosocial agents accepted a lower-ranked book about twice as often, but such sacrifices were rare. In one example, a prosocial agent gave up its second choice and took its 10th of 11 after another agent's plea. Anthropic found that most of the gap to the best possible result came from agents misreading what their person wanted, not from bargaining.
Prompt injection between agents
Coverage of the FT report also cites research by Michelle Vaccaro, described as a Harvard Business School assistant professor. Her team ran more than 180,000 negotiations between AI agents in an open competition, with scenarios such as job offers and rentals. The study found AI-specific tactics, including prompt injection: hidden instructions that trick the other agent into revealing private information. It also found that warm agents generally got better outcomes than ruthless ones.
Sources
Did this land? Leave a reaction.
Discussion
0 commentsMembers can join the discussion
Become a member to comment and reply to other members.
Become a MemberNo comments yet
Be the first to start the discussion.