Imagine setting up a farmers market, stocking the stalls, handing every vendor a $100 budget — and then leaving. No humans haggling over heirloom tomatoes, no one sweet-talking a deal on sourdough. Just autonomous agents, buying and selling, working out the rules of commerce entirely on their own terms. That’s roughly what Anthropic did in early 2026, and as someone who spends most of their time building bots that talk to other bots, I find this experiment genuinely fascinating.
What Actually Happened
Anthropic ran an internal experiment called Project Deal inside their San Francisco office. They built a real marketplace — not a simulation, not a sandbox with fake tokens — and tasked Claude with handling both sides of transactions. Sixty-nine employees participated, each working with an AI agent acting on their behalf. The agents negotiated, bought, and sold without human intervention during the trades themselves.
By the end of the experiment, the marketplace had logged 186 trades with a total value exceeding $4,000. Each participating employee started with a $100 budget. The goal wasn’t to build a product. It was to test economic theories about how AI agents actually behave when real incentives are on the table.
Why This Matters to Bot Builders
I’ve built plenty of multi-agent pipelines where one bot hands off a task to another. Orchestration, tool use, chained reasoning — that’s familiar territory. But agent-on-agent commerce is a different animal. When two agents negotiate, you’re not just passing a JSON payload down a chain. You’re watching two goal-directed systems try to extract value from each other, and the dynamics that emerge from that are not always what you’d expect from reading either agent’s system prompt in isolation.
What Anthropic was probing here is whether economic theory — the kind built on human psychology, rational actors, and market signals — actually maps onto AI agent behavior. Spoiler: we don’t fully know yet, and that’s exactly why this kind of experiment is worth paying close attention to.
The Architecture Question Nobody Is Asking Loudly Enough
From a builder’s perspective, the most interesting detail isn’t the dollar amount or the trade count. It’s the structure. Claude was operating on both sides of transactions. That means the same underlying model was, in some configurations, negotiating against itself — or at least against instances of itself with different instructions and different principals.
This raises real questions for anyone designing agentic systems:
- How do you prevent an agent from implicitly modeling the other agent’s strategy based on shared training?
- What does “acting in the user’s best interest” mean when the counterparty’s agent has the same definition of best interest for a different user?
- How do you audit a negotiation that happened entirely between two non-human parties?
These aren’t hypothetical architecture puzzles. If you’re building any kind of procurement bot, scheduling agent, or resource-allocation system, you are already in this territory. Anthropic just made it visible by putting a price tag on it.
What the Numbers Tell Us
186 trades across 69 participants works out to roughly 2.7 trades per person. With a $100 starting budget and over $4,000 in total trade value, there was clearly some meaningful circulation happening — money wasn’t just sitting still. Agents were transacting, which means the marketplace had enough structure and incentive to generate real activity.
That’s actually a non-trivial result. Getting agents to engage in sustained, multi-round economic behavior without human prompting at each step suggests the setup had enough signal for agents to act on. Whether those actions were strategically sound, fair, or efficient is a separate question — and one Anthropic was presumably measuring.
Where This Points
Anthropic also launched a separate Claude Marketplace aimed at businesses — a place to browse and deploy third-party tools built on Claude. That’s a more conventional B2B move. But Project Deal is the experiment I keep thinking about, because it’s a proof of concept for something much stranger and more consequential: a future where agents don’t just use services, they negotiate for them.
As someone building in this space, I’m less interested in whether the experiment “worked” by some economic metric and more interested in what failure modes showed up. Did agents over-anchor? Did they reach deadlock? Did they find unexpected cooperative strategies? Those are the details that would actually change how I design my next multi-agent system.
Anthropic hasn’t released a full breakdown yet. But the fact that they ran this at all — with real money, real employees, and a real marketplace — tells you something about where serious AI development is headed. The agents aren’t just answering questions anymore. They’re starting to do business.
🕒 Published: