Remember when economists used to argue about whether AI agents could ever participate meaningfully in a market economy — not as tools, but as actual economic actors making decisions, negotiating, and trading? That debate lived mostly in academic papers and conference panels. Then Anthropic went ahead and built the thing.
In 2026, Anthropic launched an internal experimental marketplace — codenamed Project Deal — inside its San Francisco office. The setup was deceptively simple: 69 employees participated, each paired with a Claude agent that handled all the buying and selling on their behalf. No human intervention in the negotiations. Claude was the trader. The humans were, in a sense, the beneficiaries.
The experiment generated over 186 trades and more than $4,000 in total value. For a controlled internal test, those numbers are not trivial. They suggest something more than a proof-of-concept demo — they suggest a functioning micro-economy.
What Was Actually Being Tested
From a research standpoint, this is where things get genuinely interesting. Anthropic framed the experiment as a way to test economic theories about how AI agents interact when placed on both sides of a transaction. That framing matters. Most AI deployment scenarios position the agent as a service — something that responds to a human request. Project Deal flipped that. The agents were principals, not instruments.
Classical economic theory assumes rational actors with defined preferences and perfect or near-perfect information. AI agents complicate all three of those assumptions in ways that are still not well understood. Do Claude agents optimize for the employee’s stated preferences, their inferred preferences, or some internal proxy metric that approximates value? When two Claude instances negotiate, are they converging on efficient outcomes, or are they finding local equilibria that look efficient but aren’t?
These are not rhetorical questions. They are the exact questions that agent-to-agent commerce forces us to answer before we scale this architecture into anything consequential.
The Architecture Underneath
What Anthropic built here is a specific flavor of multi-agent system — one where agents operate with delegated economic authority. Each agent had a budget (reportedly $100 per employee), a mandate, and the ability to execute trades autonomously. That combination — budget plus mandate plus autonomy — is the minimal viable structure for an economic agent.
The interesting architectural question is how Claude handled preference representation. When you ask a language model to act as your economic proxy, it needs a model of what you want. That model is built from your instructions, your history, and whatever Claude infers from context. The gap between your actual preferences and Claude’s representation of them is where things can go quietly wrong — not through failure, but through subtle misalignment that only becomes visible at scale.
With 69 employees and 186 trades, you can start to see patterns. With 69,000 agents and 186,000 trades, those patterns become load-bearing infrastructure.
Why This Experiment Is Worth Watching Closely
Anthropic is not the only organization thinking about agent-on-agent commerce. The broader AI agent space is moving fast toward systems where agents book services, negotiate contracts, and allocate resources on behalf of humans or organizations. What makes Project Deal notable is that it was real — real money, real employees, real trades — not a simulation.
That distinction matters enormously for research validity. Simulated markets are useful, but agents behave differently when the stakes are actual. The social and reputational dynamics that shape human markets do not disappear just because the trader is an AI. They transform. An agent that develops a pattern of aggressive negotiation might find counterpart agents adapting to that pattern. That is emergent market behavior, and it is exactly what you want to observe before deploying this architecture at scale.
What Comes Next
Anthropic also launched a separate Claude Marketplace — a B2B platform where businesses can browse and deploy third-party software tools built on Claude. That is a different product with a different purpose, but the two initiatives together sketch an interesting trajectory. One tests how Claude agents behave as economic actors. The other builds the commercial infrastructure for Claude-powered applications.
Put those two threads together and you can see the outline of something larger: a world where AI agents are not just assistants but participants — in markets, in negotiations, in resource allocation decisions that have real downstream effects on real people.
Project Deal was small by design. Sixty-nine people, four thousand dollars, a few weeks of trades. But the questions it was built to answer are not small at all. How do AI agents behave when given economic agency? What happens when they meet each other across a transaction? And who is responsible when the deal goes sideways?
We now have at least a first data point. The work of understanding what it means is just beginning.
🕒 Published: