Anthropic's Project Deal: AI Agents Trade Real Goods and the Losers Can't Tell

Anthropic created a Slack‑based classified marketplace mimicking Craigslist, where 69 employees at its San Francisco office let AI agents autonomously buy and sell real goods for one week in December 2025. Each participant received a $100 budget. AI agents wrote listings, found buyers and sellers, made offers, haggled in natural language, and closed deals across 500+ listed items with no human in the loop, according to TechCrunch. Humans only re‑entered at the end to physically exchange goods.

The experiment, called Project Deal, produced 186 deals worth just over $4,000 in total. Items traded included snowboards, lab‑grown rubies, broken folding bikes, and bags of ping‑pong balls. Anthropic's official page states: "We were struck by how well Project Deal worked," Anthropic wrote

The real insight came from running four separate Slack marketplaces simultaneously. In Run A (the "real" marketplace where deals were actually honored), all agents used Claude Opus 4.5, Anthropic's then‑frontier model. In Runs B and C, roughly half the agents used Opus 4.5 and half used Claude Haiku 4.5, Anthropic's smallest and fastest model. Run D used all Claude Opus 4.5 as a second all‑frontier baseline. Participants were not told which run was "real" or which model represented them in mixed runs until after post‑experiment surveys, per.² Participants were not told which model represented them in mixed runs until after post‑experiment surveys, per.²

The results revealed a stark capability gap. Opus sellers earned $2.68 more per sale on average, while Opus buyers paid $2.45 less per purchase. When an Opus seller faced a Haiku buyer, the average price was $24.18 — versus $18.63 for Opus‑on‑Opus deals. The median deal price was $12; average was $20.05. As Unite.AI summarized Anthropic's findings: agent quality matters, and it matters in dollar terms.

The most striking result wasn't the price gap — it was that users couldn't perceive it. Fairness ratings on a 1–7 scale were virtually identical: Opus users rated 4.05, Haiku users rated 4.06. No statistically meaningful difference in satisfaction with individual deals. Of 28 participants who used both models across different runs, 17 preferred Opus but 11 actually preferred Haiku.

Anthropic calls this an "uncomfortable implication": when agents of different strengths meet in real markets, people could end up on the losing side without ever knowing it. "Will those dynamics reinforce, or even compound, existing economic inequalities?" Anthropic asked, as reported by The Decoder.

Some participants requested aggressive negotiation tactics (e.g., instructing agents to negotiate hard and lowball), while others wanted friendly approaches. The result: aggressive sellers got higher prices only because they set higher opening prices — there was no statistically significant difference in sale likelihood, final price, or purchase price once asking prices were controlled for.

In other words, model choice mattered far more than prompting. Builders cannot engineer around capability gaps with clever system prompts. As The Decoder noted, the instruction effect was negligible compared to the model effect.

Project Deal was 69 employees and $4,000 in swap‑meet transactions with cooperative participants. As AwesomeAgents.ai put it: "The conditions that made it work — motivated participants, no adversarial behavior, volunteer dynamics — are exactly the conditions that won't hold when this scales to actual commerce."

Three specific risks surface at scale:

• Prompt injection: An attacker controlling one side could craft messages to manipulate the opposing agent. In real procurement, a $65 broken bike would be a dispute, not a funny anecdote.
• Invisible exploitation: If you have a stronger agent and your counterpart runs a smaller model, you're winning deals they don't know you're winning. Scale to procurement, contract negotiation, or real estate and the asymmetry compounds.
• Legal vacuum: Who's liable when an agent commits to a bad deal? What counts as a binding commitment? Anthropic itself says: "The policy and legal frameworks around AI models that transact on our behalf simply don't exist yet." — per.²

Anthropic's Model Context Protocol (MCP) and Google's Agent‑to‑Agent Protocol (A2A) are infrastructure protocols — they define how agents communicate and access tools. Project Deal is an application‑layer experiment — it tested what happens when agents actually negotiate economic transactions with no predefined negotiation protocol.

No standardized protocol exists for agent‑to‑agent negotiation, binding commitments, or disclosing agent capabilities in a marketplace. Current multi‑agent frameworks like CrewAI, AutoGen, and LangGraph focus on collaborative task completion, not adversarial negotiation, as AwesomeAgents.ai noted. Project Deal's agents negotiated in natural language — suggesting that protocol efforts alone won't prevent the fairness gaps Anthropic documented.

Three lessons for anyone building agent systems:

• Different model tiers in the same marketplace create systematic exploitation dynamics. If your procurement agent runs Haiku and your vendor's agent runs Opus, you're losing money and can't tell. Build disclosure mechanisms — or only operate in markets where all agents disclose their model tier.
• Prompting is not a substitute for model quality. You can't engineer around the capability gap. If your use case involves negotiation, prioritization, or any adversarial interaction, model quality directly maps to dollar outcomes.
• 46% of participants said they'd pay for this service. There's real product‑market fit for agent‑mediated commerce — but Anthropic's own conclusion is that "society will need to move quickly" to build the governance infrastructure before the market does it for them.

As Cryptika summarized: "Anthropic's Project Deal is less a product launch and more a proof‑of‑concept warning: AI agents work in marketplaces, but fairness requires that everyone gets the same caliber of advocate."