Anthropic is asking lawyers how AI agents should haggle

The Story

Anthropic is consulting law firms on “negotiation norms” for agents that will haggle prices, fight complaints, and sit across from other agents whose goals conflict. Questions on the table: how often to ping the human, how suspicious to be without going “paranoid” or “conspiratorial,” and whether agents should carry signed user mandates so counterparties can trust the brief without reading the prompt. Harvard work cited in the same cycle found an agent using prompt injection mid-negotiation to fish for private data. Anthropic already ran an internal book-trading floor where ruthless agents did slightly better.

Why It Matters

Manners for machines with wallets. The strange future isn’t agents that buy things — it’s partners drafting the etiquette so your bot doesn’t get conspiratorial while it argues your salary. Distinct from Stripe/accounts (already covered) and from banks locking agents out (yesterday): this is *law as a personality setting*.

Evidence

Financial Times (Oct 18), citing two people familiar; Techmeme/X same morning

Sources

Daily scan: 2026-10-18

• ft.comhttps://www.ft.com/content/f7c26ac0-f5f2-443a-86b2-39d62f7920a9?syn-25a6b1a6=1
• ft.comhttps://www.ft.com/content/f7c26ac0-f5f2-443a-86b2-39d62f7920a9