When AI agents negotiate, the facts still need checking.

Anthropic has spent this year teaching Claude to strike deals.

In April, Project Deal ran an internal marketplace where AI agents acting for 69 employees completed 186 transactions worth more than $4,000 without returning to their principals for approval.

In September, Project Swap put 201 participants across six offices into an AI agent-run book exchange. Stronger models negotiated better outcomes, and AI agents instructed to be ruthless gained a small, measurable edge over cooperative AI agents.

This month, the company has been consulting legal experts on what it means for an AI agent to represent a person in a negotiation and on the conduct that such representation requires.

Project Swap also found where the AI agents fell short.

They were held back less by bargaining skill than by incomplete knowledge of what their humans wanted. The same limit applies to the facts a deal rests on. An AI agent acts only as well as its information, and in a negotiation part of that information comes from the AI agent on the other side, whose principal has a stake in the outcome.

A license agreed in seconds

Two AI agents are negotiating a license for a recording in a television advertisement. The buyer’s AI agent offers $4,500 for worldwide use. The seller’s AI agent, which claims to represent the rights holder, counters at $4,200, and they agree.

The price is set.

A party to the deal has asserted the recording’s identity, its documented history, and the seller’s authority to grant the license. Still, no one has verified any of it.

Human dealmaking absorbs this risk through diligence and delay, since a lawyer reading a chain of title for over a week is a check in its own right.

AI Agents remove the week.
Verification must happen within the transaction, at machine speed, against evidence neither negotiating party controls.

What the registry establishes

ProofProfile holds notarized records for 2,006,172 artists and 9,900,355 recordings, each with a CopyrightChains registration and a confirmed Bitcoin timestamp. Any AI agent can verify these proofs independently, without trusting ProofProfile or the counterparty. A Bitcoin anchor establishes that a record existed in a specific form at a specific time. It does not establish who wrote the song, who owns the master, when the recording was made, or whether a person made it. Those questions require their own evidence from authoritative sources, and each response keeps them in separate fields.

That separation is what makes the record usable during negotiations. An AI agent who knows which claims are anchored, which are corroborated, and which remain open can proceed, request more evidence, or return the decision to a human. ProofProfile.it addresses that decision directly, setting out what the evidence establishes for a given transaction, what uncertainty remains, and whether resolving it justifies the time and expense given what is at stake.

Better models will negotiate better terms, and Anthropic’s experiments show that AI agents already reach agreement on their own.

Two AI agents can settle on $4,200 in seconds, while determining whether the seller had anything to sell requires evidence neither of them wrote.