Anthropic Sends the Agent to Contract School

Anthropic is turning to lawyers to help AI agents strike deals, the Financial Times reports. The work concerns machines negotiating on humans’ behalf. The available report summary does not identify participating lawyers, governing jurisdictions, or particular transactions. It establishes a legal-design problem, not a completed system for enforceable machine-made agreements.

The mechanism is delegation. A person or business gives software instructions, and the software communicates with another party. Three questions follow: What may the agent propose? What may it accept? What still requires human approval? An agent can produce a convincing acceptance without establishing that its operator authorized the commitment. Fluency does not supply a signature policy.

Useful delegation is possible without granting unlimited discretion. A business could authorize negotiation within a price range while reserving changes to delivery obligations for a manager. That is an illustrative design, not a documented Anthropic feature. The important distinction is between preparing an offer and committing the business to it. A demonstration that secures a discount tests neither distinction by itself.

Platform rules add another layer. Anthropic can set operating limits for its service. Customers need to define their own commercial mandates within those limits. The test is whether customers can inspect and enforce those mandates, rather than merely express preferences that the software interprets. A provider’s safeguard might prevent a prohibited transaction while doing nothing to prevent an authorized transaction on commercially unacceptable terms. Those are different protections for different interests.

The counterparty also needs an account of authority. A negotiating business should be able to establish whom the agent represents and whether an apparent acceptance requires further approval. If a disputed commitment follows, the relevant questions include the instructions in force, the messages exchanged, and any human confirmation. Applicable law and the transaction’s terms would shape the outcome. Calling the incident an agent error would not settle which organization must answer for it.

Anthropic and participating businesses can specify these boundaries before celebrating negotiation performance. They can state who grants authority, which terms require confirmation, how revocation reaches an ongoing negotiation, and what records each party receives. They can also name the procedure for challenging an unauthorized commitment and the organization responsible for responding. The next useful demonstration would show that process working when a customer says no.