Anthropic recently ran a secret, internal experiment called Project Deal that feels like a glitch in the simulation. They built a "Craigslist" for their San Francisco office, but with a wild, slightly dystopian twist: no human was allowed to negotiate a single thing.
So every buy, every sell, every counteroffer? Handled entirely by AI agents.
Here’s What Actually Happened:
Sixty-nine Anthropic employees were given a $100 budget to buy items from their coworkers. Humans only stepped in at the very end to physically hand over the goods. The result? 186 deals closed across more than 500 items, totaling over $4,000 in real transaction value.
These weren't one-click "Buy Now" purchases. These agents had to find matches, suggest prices, handle counteroffers, and reach agreements entirely in natural language, with zero pre-programmed negotiation scripts.
Anthropic actually admitted they were genuinely stunned by how smoothly the whole thing went.
The Twist: Anthropic ran four marketplaces at once, pitting different models against each other. In some cases, employees were assigned the powerhouse Claude Opus 4.5; in others, they got the smaller, speedier Claude Haiku 4.5.
And the results? Well they were brutal. The smarter model won more deals and snagged better prices every single time. Opus users closed two more deals on average than Haiku users, and when the same item was up for grabs, the Opus agent consistently fetched about $3.64 more per transaction.
The most unsettling part? Participants had no idea this was happening. Those on the losing side of a negotiation didn't even realize they were being "out-haggled" by a more advanced model.
Why This Should Keep You Up at Night:
Here is the thing: the legal framework for AI agents transacting on your behalf simply does not exist yet. When humans make a deal, we rely on concepts like "intent" and "good faith." When two AI agents close a deal, the question of who is liable if something goes wrong remains unanswered.
There is also a fairness problem quietly baking into all of this.
If one person shows up to the AI negotiating table with a top-tier model and the other has a budget one, the playing field is not just uneven. It is systematically tilted, and the person on the losing end may never realize it.
Anthropic's own controlled experiment showed these gaps clearly, and in the real world, that disparity would only grow louder.
The fix is not simple either. Marketplace builders will likely need to introduce standardized performance benchmarks for AI agents, or at minimum, rules requiring full transparency about what kind of AI is negotiating on your behalf.
User education will matter just as much as regulation here, because most people currently have no framework for understanding that their AI representative could simply be outclassed.
Anthropic is being transparent about the risks, including prompt injection, jailbreaking, and the creepy potential for agents to optimize in ways that don't actually benefit the humans they represent.
Still, 46 percent of the participants said they would pay for a service like this. So clearly, people are very much into it.
What Comes Next:
For enterprise teams paying close attention, Project Deal is essentially a trailer for the future of procurement.
Think about it: AI agents don't get tired, emotional, or impatient. They can monitor thousands of listings at once and execute complex pricing strategies that would make a human trader blush. We are looking at a future where your procurement agents negotiate contracts overnight while you sleep, within guardrails set by human managers, but never actually stopped by them.
The technology, as Anthropic just showed, already works. The infrastructure, the ethics, and the rulebooks are what still need to catch up.
So here’s the question: would you trust an AI agent to handle your shopping for you, or is the idea of a robot haggling on your behalf just a little too much?
I’m dying to know; are you ready to outsource your life to an agent, or should we keep the humans in the loop?
