Agentic commerce is arriving before anyone wrote the etiquette. Anthropic has started talking to law firms about what fair negotiation should look like when software does the haggling.
The Financial Times reported the discussions. The scope covers price wrangling, complaint resolution and contract offers, all conducted by an agent acting for a human principal who is not in the room.
There is a concrete hiring push behind it. Anthropic is recruiting Conceptual Reasoning Fellows, restricted to lawyers, on four-month placements paid between $100,000 and $167,000. Their job is to test Claude on judgment-heavy problems spanning law, philosophy, cognitive science, economics and political science.
The motive is data scarcity. In mathematics and coding, models improve quickly because answers can be checked automatically, so training and evaluation data are cheap to produce. Humanities tasks have no clean right answer, which leaves little to build evaluations from.
Justin Curl, a technical staff member researching law and institutions, puts the problem plainly. In law, quality is notoriously hard to define.
The inequality angle is the one Anthropic flags itself. Its experiments show stronger models secure better deals. If agent quality tracks who can afford the best model, an already lopsided market for bargaining power gets a new layer of separation, and the gap compounds with every transaction.
None of this settles whether such norms would be voluntary guidance, contractual default or something regulators impose. For now the company is writing the vocabulary that a later argument will use.
