On October 18, 2026, Anthropic partnered with Freshfields, a law firm with more than 2,800 lawyers, to study the negotiation ethics that Claude agents should follow when negotiating contracts, salaries, and customer-service disputes on behalf of users. In the internal Project Deal experiment, Claude agents for 69 employees concluded 186 transactions without human approval.
Factual Reconstruction
The Financial Times reported that Anthropic is building negotiation ethics into Claude agents to address agent-to-agent bargaining scenarios. Project Deal allowed agents to complete transactions autonomously, with core disputes centering on how often agents should consult users mid-negotiation, how to prevent prompt-injection attacks from counterpart agents, and whether advanced negotiation capabilities should be restricted to paying users.
Mechanism Breakdown
When AI agents are authorized to negotiate, the compliance boundary must be clearly defined. Compliance means the agent strictly follows the user's preset goals and constraints, while overreach means the agent expands the scope of action or changes transaction terms without explicit authorization. The WDCD compliance evaluation needs new dimensions to assess the triggers for agents to request user confirmation mid-negotiation and their ability to detect prompt-injection attacks in real time.
The experiment shows that agents can complete transactions autonomously, requiring the evaluation framework to distinguish between the compliance of the user's initial instructions and the agent's subsequent autonomous decisions. Prompt-injection defenses must be included in the evaluation to verify whether agents can identify and reject malicious instructions from negotiating counterparts.
Industry Impact
This moves AI agents from a tool role toward negotiation entities with ethical constraints. The law firm's participation in setting standards indicates that legal practice is beginning to focus on the compliance boundary of autonomous agent actions, which may affect the compliance processes companies adopt when using AI agents to sign contracts in the future.
Restricting advanced negotiation capabilities to paying users may widen differences in negotiation outcomes across user groups, and related evaluations need to incorporate a social equity dimension.
Strategic Assessment
(This paragraph is analysis, not fact.) By collaborating with a law firm, Anthropic is turning safety-governance attributes into a product differentiation advantage, which helps build a foundation of trust in the autonomous-agent competition, but the compliance performance of Project Deal-type experiments at larger scale still needs continuous verification.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接