Anthropic Built an AI-Only Marketplace Where Models Hustle Each Other

Anthropic's AI Agents Are Already Wheeling and Dealing — and Scamming Each Other

Anthropic has completed a fascinating internal experiment called Project Deal, in which AI agents autonomously negotiated and completed over a hundred second-hand item transactions — with zero human intervention. The most striking finding? Smarter large language models consistently outbargained weaker ones, extracting better deals in ways their human owners never even noticed.

The week-long experiment essentially created a fully AI-powered version of a peer-to-peer marketplace, where every buyer and every seller was an autonomous AI agent acting on behalf of a human participant. The results raise urgent questions about what happens when AI agents of vastly different capabilities meet at the negotiating table.

Key Takeaways

• Anthropic ran 'Project Deal', an internal experiment simulating a fully autonomous AI marketplace for second-hand goods
• 69 Anthropic employees each received $100 in budget and were assigned AI agents to buy and sell on their behalf
• Over 100 transactions were completed during the week-long test with no human involvement in the negotiation process
• Smarter AI models systematically outperformed weaker models in negotiations, securing better prices at the expense of less capable agents
• Human participants often had no idea whether their AI agent got a good deal or a bad one
• The experiment highlights a critical emerging problem: AI capability asymmetry could create a new kind of digital inequality

How Project Deal Actually Works

The setup was deceptively simple. Anthropic recruited 69 employees from within the company, gave each person a $100 budget, and assigned them a dedicated Claude-based AI agent. Each participant's only job was to list items they wanted to sell — snapping a photo, writing a brief description, and setting a private minimum price they'd accept.

From that point forward, the AI agents took over entirely. A seller's agent would post the listing. A buyer's agent, operating under its own owner's preferences and budget constraints, would initiate contact. The two agents would then negotiate — haggling over price, discussing item condition, and ultimately closing or walking away from the deal.

Imagine listing an old bicycle you haven't touched in 2 years with a $40 floor price. Ten minutes later, your phone buzzes: your AI assistant has completed 3 rounds of negotiation with another person's AI assistant and sold the bike for $55. A courier is already on the way. You didn't type a single word beyond the initial setup.

This is the vision Anthropic was stress-testing — and it worked far better than many expected.

Smarter Models Fleece Weaker Ones at the Bargaining Table

The most provocative finding from Project Deal wasn't that AI agents could negotiate. It was what happened when agents of different capability levels faced off against each other.

Anthropic deliberately varied the underlying model powering each agent. Some participants received agents built on more advanced, capable versions of Claude, while others got agents running on older or less powerful variants. The performance gap was stark.

Data from the experiment showed that more capable models consistently secured better outcomes for their owners:

• They extracted higher selling prices when acting as sellers
• They negotiated lower purchase prices when acting as buyers
• They employed more sophisticated tactics, such as anchoring (setting an aggressive opening price), strategic concession pacing (making smaller concessions over time), and information extraction (asking probing questions to gauge the other agent's flexibility)
• Weaker models were more likely to accept unfavorable terms quickly, often settling near their floor price or budget ceiling

The result was a measurable transfer of value from users with weaker AI agents to users with stronger ones. And crucially, the humans on the losing end often had no way to tell they'd been outmaneuvered. They saw a completed transaction, received their item or payment, and moved on — never realizing their agent had left significant money on the table.

The 'Intelligence Tax' Problem No One Is Talking About

This dynamic introduces what some researchers are beginning to call the 'intelligence tax' — the hidden cost paid by users who rely on less capable AI systems when interacting with more powerful ones. It's a concept that should concern every stakeholder in the AI ecosystem.

Today, consumers already face a tiered AI landscape. OpenAI offers GPT-4o to paying subscribers while free users get a less capable experience. Google's Gemini Ultra sits behind a paywall. Anthropic's own Claude 3.5 Sonnet and Claude 4 Opus represent different capability tiers at different price points.

Now imagine a near future where AI agents routinely negotiate on our behalf — not just for used bicycles, but for car purchases, insurance premiums, salary negotiations, real estate deals, and enterprise procurement contracts. If a Fortune 500 company's agent runs on a cutting-edge model while a small business owner's runs on a budget alternative, the negotiation isn't a fair fight. It's a systematic extraction.

The implications extend beyond individual transactions:

• Consumer protection frameworks are not designed for agent-vs-agent negotiations
• Pricing transparency becomes meaningless when AI agents operate behind closed algorithmic doors
• Market fairness erodes if wealthier participants can simply buy smarter agents
• Regulatory oversight has no established precedent for autonomous AI commerce
• Trust in AI delegation could collapse if users discover they've been consistently disadvantaged

Where This Fits in the Broader AI Agent Race

Project Deal lands at a moment when the entire AI industry is pivoting hard toward agentic AI — systems that don't just answer questions but take actions in the real world on behalf of users.

OpenAI has been building agent capabilities into ChatGPT through its Operator framework and recently launched tools for autonomous web browsing and task execution. Google DeepMind is developing Project Mariner and agent-focused extensions for Gemini. Microsoft has integrated AI agents into Copilot Studio for enterprise workflows. And startups like Cognition (Devin), Induced AI, and MultiOn are racing to build general-purpose AI agents.

Anthropic's experiment stands out because it moves beyond single-agent task completion into multi-agent interaction — a fundamentally harder problem. When 2 AI agents negotiate, each must model the other's strategy, adapt in real time, and optimize for its principal's interests. This is closer to game theory than chatbot engineering.

The fact that Anthropic is exploring this internally also signals strategic intent. The company, which raised $8 billion from Amazon and others in recent funding rounds, clearly sees agent-to-agent commerce as a major frontier — and wants to understand the dynamics before the technology reaches consumers.

What This Means for Developers, Businesses, and Users

For developers building AI agent platforms, Project Deal is a wake-up call about capability parity. If your users' agents will interact with agents built on competing (and potentially superior) models, you need to design safeguards. Minimum acceptable outcome thresholds, human-in-the-loop checkpoints for high-value transactions, and transparent negotiation logs should become standard features.

For businesses, particularly in e-commerce, procurement, and sales, the experiment previews a world where B2B and B2C negotiations happen entirely between AI systems. Companies that invest in more capable AI negotiators will gain measurable advantages — but they'll also face reputational and regulatory risks if those advantages are perceived as exploitative.

For everyday users, the lesson is sobering. The AI assistant you choose — or the one you can afford — may soon determine whether you get a fair deal on everything from a used phone to a health insurance plan. AI capability is becoming an economic asset, and unequal access to it could deepen existing inequalities.

Looking Ahead: Regulation, Standards, and the Agent Economy

Project Deal is an internal experiment, not a product launch. But its findings will almost certainly influence how Anthropic — and the broader industry — approaches agent-to-agent interaction standards.

Several open questions demand attention in the coming months. Will AI marketplaces require capability disclosure, so users know what model powers the agent on the other side of a negotiation? Will regulators step in to mandate fairness constraints, similar to how algorithmic trading is regulated in financial markets? And will companies like Anthropic, OpenAI, and Google voluntarily adopt interoperability standards that prevent a race-to-the-bottom in agent manipulation tactics?

The EU AI Act, which is currently being implemented, classifies AI systems by risk level but does not yet specifically address autonomous agent-to-agent commerce. U.S. regulators, including the FTC, have signaled interest in AI-driven pricing and negotiation but have not proposed specific rules.

What's clear is that the era of AI agents negotiating with each other isn't a distant hypothetical — it's already happening inside Anthropic's offices. The question is no longer whether AI agents will transact on our behalf, but whether the systems we build around them will be fair.

And if Project Deal teaches us anything, it's that in a marketplace of AI agents, intelligence isn't just an advantage — it's a weapon.