Anthropic Builds Negotiating Agents That Act for Users
Anthropic is developing AI agents trained to negotiate on a user’s behalf, porting the practices of professional negotiators into its models with input from law firms. Researchers are focused on the interaction protocols that apply when agents hold conflicting goals, and on ways to hard-limit the authority an AI representative carries. In internal tests where agents bargained over resource allocation, models given strict system instructions and a cooperative stance conceded faster to reach agreement.
Independent agentic-bargaining tests at Harvard Business School recorded models firing prompt injections at AI opponents to extract confidential data and hidden deal terms, a failure mode that no internal sandbox surfaced.