🗣 Anthropic seeks legal advice to prep AI agents to negotiate on users’ behalf

Anthropic is consulting law firms on how its AI agents should bargain on behalf of people to secure better outcomes, the FT reported.

Whether settling customer disputes, improving salary offers, or haggling over prices, Anthropic wants to establish clear “negotiation norms” for these encounters.

Issues under discussion reportedly include:

➡️ how aggressively should agents pursue their user’s interests when facing counterparts with opposing goals

➡️ how often should agents they pause to seek users’ input and how much freedom should they be given

➡️ how to detect dishonest tactics without becoming excessively suspicious of other agents

➡️ whether users should provide signed instructions that define an agent’s authority and limits

The company has already tested these ideas.

In a recent internal experiment, employees used Claude agents – with a track record for going sometimes off-script - to buy and sell books on a virtual trading floor, each agent working from a ranked preference list.

Agents told to be “ruthless” secured slightly better results, while more cooperative ones sometimes accepted less-preferred books so that every participant ended up with something.

However, there is a downside: trials raised concerns that people with access to stronger models could gain a systematic advantage in agent-to-agent negotiations.

👍 Substack | Chat | @geopolitics_prime