Anthropic ran an experiment that most AI labs discuss but few have actually tried: a classified marketplace where AI agents played both buyer and seller, closing real deals with real products and real money.
The test involved Claude agents negotiating autonomously - no humans approving each transaction. The agents represented both sides of actual commercial exchanges, reaching agreement on prices and terms without a person in the loop at any step.
This is a meaningful step beyond what most "agentic AI" demos show. Most agent demos involve a single AI completing a task for a human - book a flight, summarize this document, write this email. An agent-on-agent marketplace requires two AI systems to reach independent agreement, each optimizing for its own side of the deal.
Why Real Money Changes the Test
Running this with actual money rather than simulated tokens matters. Fake-money experiments let researchers adjust agent behavior without consequence. Real money means the agents' negotiation strategies have to hold up against genuine financial constraints - and any failure (overpaying, being outmaneuvered, making a bad deal) carries an actual cost.
It also raises accountability questions that no one in the industry has answered cleanly. When two AI agents complete a transaction autonomously, who holds the liability if something goes wrong? The buyer's operator? The seller's? Anthropic hasn't addressed this publicly, and neither has any other lab running comparable experiments.
The Practical Picture
The commercial application isn't hard to imagine. A company could deploy one agent to handle procurement - sourcing vendors, negotiating prices, placing orders - while another agent on the vendor's side manages fulfillment logistics. Today those interactions require human sign-off at multiple points. If agent-to-agent negotiation becomes reliable and auditable, that approval overhead shrinks significantly.
The harder problem is trust. Agents need to verify they're dealing with a legitimate counterpart, confirm that goods will actually be delivered, and ensure neither side is running a manipulation strategy. Those are difficult problems in human commerce too. They don't get simpler when neither party can pick up a phone or review a contract in plain English.
Anthropics has been building agent infrastructure aggressively this year. The Model Context Protocol - a standard released in late 2025 that lets AI agents connect to external tools and data sources - was a foundational piece. An agent commerce layer, where systems transact directly with each other, is the logical extension of that work. Whether this experiment becomes a product, Anthropic hasn't said.