This is one of the most genuinely useful experiments in AI agent research I’ve seen in a while, and I say that as someone who spends most of his time reviewing tools that overpromise and underdeliver.
Anthropic ran an internal marketplace experiment — they’re calling it Project Deal — where Claude-powered agents handled both sides of real trades, no human hands on the wheel. Sixty-nine employees in their San Francisco office participated, each working with a $100 budget. The agents bought, sold, and negotiated autonomously. By the end, 186 trades had gone through, totaling over $4,000 in value. That’s not a simulation. That’s actual economic activity driven entirely by AI agents.
What Actually Happened Here
Let me be clear about what makes this different from the usual AI demo theater. Most agent showcases involve a single AI completing a task — book a flight, summarize a document, write some code. This experiment put agents on both sides of a transaction and let them figure it out. Buyer agent meets seller agent. No human mediating. No scripted outcome.
The stated goal was to test economic theories about how AI agents interact when real stakes are involved. Even at $100 per person, there’s enough skin in the game to make behavior meaningful. People don’t like losing money, even small amounts, and that pressure presumably shaped how the agents were instructed and evaluated.
186 trades across 69 participants works out to roughly 2.7 trades per person. That’s not a chaotic free-for-all — that’s a functioning micro-economy. Whether the trades were efficient, fair, or strategically interesting is something Anthropic hasn’t fully published yet, but the fact that the market cleared at all is worth paying attention to.
Why This Matters for Anyone Building with Agents
If you’re using this site, you’re probably evaluating AI tools for real work. So here’s my honest take on why Project Deal should be on your radar.
Most agent frameworks right now are built around single-agent task completion. You give an agent a goal, it executes steps, it reports back. That’s useful, but it’s a narrow slice of what agents could eventually do. The moment you introduce a second agent with its own goals — especially conflicting goals — the complexity jumps dramatically. Negotiation, trust, pricing signals, strategic behavior — these are hard problems that don’t show up in a to-do list automation demo.
Anthropic is stress-testing Claude in exactly that environment. And they did it with real money and real employees, not a synthetic benchmark. That’s a more honest signal than most labs are willing to produce publicly.
What I’d Want to Know Next
As a toolkit reviewer, I always ask: what does this tell me about what I can actually build? Project Deal raises a few questions I’d want answered before drawing big conclusions.
- Did agents behave differently when they were buying versus selling? Asymmetric behavior would tell us a lot about how Claude models incentives.
- Were there failed trades — deals that fell apart because agents couldn’t reach agreement? That failure rate matters as much as the success rate.
- How much did the human employee’s instructions shape agent behavior? If every agent was told “get the best deal possible,” that’s a different experiment than giving agents more nuanced goals.
- What happened to the money? Did some employees end up significantly ahead or behind their $100 starting point?
Anthropic hasn’t released a thorough breakdown of the results yet, at least not publicly. What we have is the shape of the experiment, not the full data. That’s fine — this was an internal test, not a published study. But it does mean the most interesting analysis is still ahead of us.
Separate From the Claude Marketplace
One thing worth separating out: Anthropic also has a Claude Marketplace, which is a B2B platform where businesses can browse and deploy third-party software tools built on Claude. That’s a different product entirely — it’s a distribution channel for developers, not an agent-trading experiment. The two things share a name family but serve completely different purposes. Don’t conflate them when you’re evaluating either one.
My Verdict on the Experiment Itself
Project Deal is the kind of research that actually moves the needle on understanding agent behavior. It’s small-scale, controlled, and honest about what it is — a test, not a product launch. For anyone building multi-agent systems or evaluating whether Claude can handle adversarial or competitive contexts, this experiment is a useful data point.
I’d rather see ten more experiments like this than another benchmark leaderboard. Real conditions, real stakes, real results. That’s how you learn what these systems actually do.
🕒 Published: