Anthropic Seeks Legal Advice on Rules for AI Agents Negotiating for Users
The AMW Read
Legal consultation incrementally updates Anthropic's agent strategy, while delegation limits and preference understanding are deployment constraints across the agent market.
Anthropic Seeks Legal Advice on Rules for AI Agents Negotiating for Users
Anthropic is consulting law firms on how AI agents should negotiate on users' behalf, Platum reported, citing an October 18 Financial Times report based on two people familiar with the discussions. Questions include how strongly agents should prioritize their users, when they should seek approval and how skeptically they should treat counterparties. A signed delegation document establishing an agent's authority and constraints is among the options being considered.
The discussions expose a deployment barrier for autonomous agents: negotiating successfully requires understanding a user's preferences and staying within delegated authority. In Anthropic's Project Deal experiment, 69 employees delegated secondhand trading to Claude agents, which completed 186 transactions. Agents using Opus 4.5 obtained better prices than those using Haiku 4.5, while participants using the weaker model did not clearly recognize their disadvantage. In the subsequent Project Swap book exchange, Anthropic attributed 85% of the gap between optimal allocation and actual outcomes to agents' incomplete understanding of user preferences. These internal experiments identify risks; they do not establish reliability in commercial negotiations.
For builders, the concrete implication is to evaluate preference understanding and approval boundaries alongside transaction outcomes. A system can secure an agreement while accepting terms its user would reject. For investors, legal consultation signals work on a commercialization constraint, rather than proof that it has been solved: the report describes discussions and proposed safeguards, without establishing an adopted standard for authority or liability.
