🤖 Anthropic is preparing for a world where AI agents negotiate against each...

🤖 Anthropic is preparing for a world where AI agents negotiate against each other Anthropic is reportedly consulting law firms on how Claude agents should handle these negotiations, including when they can make decisions independently and when they must ask humans for approval. The company is also exploring digitally signed documents to verify that agents are following their users’ actual instructions. But early experiments reveal a potential problem: AI agents can get outplayed without their users even realizing it. In one experiment, Claude Opus agents sold items for $2.68 more and bought them for $2.45 less than Claude Haiku agents. Yet users of the weaker models rated their deals almost identically in fairness: 4.06 versus 4.05 out of 7. Another experiment involving 201 Anthropic employees found that instructing agents to be “ruthless” barely improved their results. Upgrading from Haiku to Opus had a much larger effect on negotiation performance. The challenge goes beyond making agents better negotiators. They must also know when to trust competing agents, detect suspicious behavior, and avoid becoming so distrustful that every interaction looks like a scam. @aipost 🏴


