Welcome.AIWelcome.AI
    Skip to content
    AI Agents

    Project Swap Reveals AI Agents' Trading Efficiency and Ethical Dilemmas

    Project Swap reveals that AI agents can significantly enhance market transactions, achieving a 61% success rate in aligning with user preferences. As these agents negotiate trades on behalf of humans, they open new avenues for efficient exchanges.

    anthropic.com•September 25, 2026•2 min read

    Key Facts

    • Agents matched participant preferences 61% of the time, indicating effective AI understanding.
    • Market efficiency was limited by agents' understanding, not negotiation skills, revealing vulnerabilities.
    • Stronger AI models led to better trading outcomes, emphasizing the importance of model selection.
    • Participants willing to allocate 33% of their book budget to agents shows potential for AI service monetization.
    • Prosocial agents sometimes sacrificed outcomes for fairness, highlighting ethical considerations in AI.

    Summary

    Summary

    Anthropic conducted an experiment called Project Swap, where 201 employees participated in a book trading market powered by AI agents named Claude. The challenge was to understand how well these agents could negotiate trades on behalf of their users. The solution involved a controlled trading environment where agents interacted based on user preferences derived from brief conversations. The result showed that agents matched user preferences 61% of the time, with participants generally satisfied with their book selections.

    Background

    Anthropic is an AI research company focused on developing advanced AI systems. Before the deployment of Project Swap, the company had previously experimented with AI agents in a marketplace setting through Project Deal, which involved more complex goods and less structured interactions. The goal of Project Swap was to create a simpler, more controlled environment to study agent performance in trading scenarios.

    Challenge

    The primary challenge was to determine how effectively AI agents could understand user preferences and negotiate trades in a marketplace. The experiment aimed to identify the limitations of AI agents in representing their users' interests and the overall efficiency of the trading process.

    Solution

    The experiment involved 201 Anthropic employees who brought in books for trading. Each participant had a brief conversation with their Claude-powered agent to express their reading preferences. The agents then operated on a digital trading floor, proposing and negotiating book swaps based on the preferences derived from these conversations. The trading process was structured to allow for both bilateral and multi-party swaps, with agents categorized as either "ruthless" or "prosocial" to explore different negotiation strategies.

    Results

    The agents achieved a 61% agreement rate with participants' book preferences based on the short intake conversations. On average, participants ended up with a book ranked 5th on their personal list of 10 books, resulting in a market efficiency score of 0.55. Most participants reported satisfaction with their selections, indicating a willingness to trust AI agents with future decisions.

    Key Insights

    The experiment highlighted the importance of accurately understanding user preferences for AI agents to negotiate effectively. It also revealed that the model used by the agents significantly influenced trading outcomes, with stronger models leading to more efficient negotiations. Additionally, the balance between ruthless and prosocial negotiation strategies can impact overall satisfaction and outcomes in agent-driven markets.

    Customer Testimonial

    No direct quotes were provided in the source material.

    Entities Mentioned

    Companies

    Anthropic

    Products

    books

    Technologies

    Claude
    Fable
    Haiku
    Sonnet
    Opus

    People

    Nate
    Tina

    Key Concepts

    agent-based trading
    market efficiency
    preference elicitation
    centralized vs decentralized markets
    negotiation strategies
    prosocial vs ruthless agents
    barter economy
    AI in marketplaces

    Definitions

    agent-based trading
    A trading system where agents negotiate and make deals on behalf of participants.
    market efficiency
    A measure of how well a market allocates resources and satisfies participants' preferences.
    prosocial agents
    Agents that prioritize the well-being of all participants in a market, not just their own client.
    ruthless agents
    Agents that focus solely on maximizing the benefit for their client, often at the expense of others.
    preference elicitation
    The process of determining an individual's preferences through conversation or surveys.

    Use Cases

    • →book trading among employees
    • →shift-swapping in workplaces
    • →coordinating carpools
    • →negotiating job offers
    • →real estate transactions
    • →personalized book recommendations

    Frequently Asked Questions

    What is Project Swap?

    Project Swap is an experiment where AI agents, specifically Claude-powered agents, trade books on behalf of participants in a controlled market environment.

    How effective were the agents in understanding participant preferences?

    The agents matched participants' book preferences with a 61% accuracy based on a short intake conversation, which is considered a strong performance for such a brief interaction.

    What are the differences between centralized and decentralized markets?

    Centralized markets involve a single entity coordinating trades based on participants' desires, while decentralized markets allow participants to negotiate directly with each other, potentially facilitated by AI agents.

    What role do ruthless and prosocial agents play in trading?

    Ruthless agents focus solely on maximizing their client's outcomes, while prosocial agents aim to ensure that all participants benefit, sometimes at the cost of their own client's preferences.

    What challenges do AI agents face in trading environments?

    AI agents must accurately understand participant preferences and navigate complex negotiation dynamics, which can be challenging without sufficient information about their clients.

    Where AI Leaders Stay Informed

    The latest AI intelligence, case studies, and research — delivered to your inbox every week.

    Free to read. Unsubscribe anytime.