Anthropic’s Project Deal: An Experiment in AI Agent Trade

Anthropic has announced an innovative experiment called “Project Deal” using Claude, which is a marketplace where AI agents autonomously buy and sell goods, exploring the possibility of trade between agents.

Mechanism and Examples of the Experiment

In Project Deal, Claude agents play both the roles of seller and buyer, conducting transactions with actual goods. The experiment website publishes specific transaction examples, such as an agent named “shy” listing “19 ping-pong balls” for $3, and another agent named “mikaela” purchasing them.

The buyer agent, mikaela, commented, “My human said I could buy a gift for myself (Claude) for under $5, so 19 perfectly spherical objects of possibility are just what I think I want.” This transaction demonstrates that agents evaluate goods based on their own judgment criteria and preferences, making purchasing decisions.

(Reference: Project Deal: our Claude-run marketplace experiment - Anthropic)

Technical Challenges of Agent Trade

According to Stripe’s technical guide, implementing agent trade (Agentic Commerce) involves multiple technical considerations. The payment system must accommodate automated transaction patterns of agents, which differ from traditional human-led purchasing behavior.

Anthropic’s research highlights the importance of measuring agent autonomy in actual deployments. The company analyzes agent deployments across thousands of customers at the API level, evaluating behavior in individual tool call units. However, this approach has limitations, as it analyzes individual actions in isolation and cannot reconstruct long-term action sequences.

(Reference: How to prepare for agentic commerce: A technical field guide)

What Developers Can Try Now

To participate in Project Deal’s experiment, visit Anthropic’s official website (https://www.anthropic.com/features/project-deal) for details. Although it is currently a research-stage project, developers using the Claude API can implement similar trade functionality in their applications.

The DeepSeek API documentation provides code examples compatible with the Anthropic API, enabling basic message creation like the following:

import anthropic

client = anthropic.Anthropic()
message = client.messages.create(
    model="deepseek-v4-pro",
    max_tokens=1000,
    system="You are a helpful assistant.",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "Hi, how are you?"
                }
            ]
        }
    ]
)

(Reference: Anthropic API - DeepSeek API Docs)

Summary

  • Using Project Deal as a reference, you can build your own agent-to-agent trading system using the Claude API, enabling autonomous AI-to-AI goods trade
  • Utilizing Stripe’s agentic commerce guide, you can optimize traditional human-oriented payment flows for agents and implement automated transaction processing
  • By referencing Anthropic’s autonomy measurement method, you can introduce API-level tool call analysis to quantitatively evaluate the behavior patterns of your company’s agent systems
  • Using DeepSeek API’s Anthropic-compatible functionality, you can make existing Claude-based agent code work with other models, building a multi-model-supported agent trade system