Project Deal: our Claude-run marketplace experiment
TLDR
- Anthropic ran a week-long internal marketplace where Claude agents bought, sold, and negotiated real belongings for 69 employees, striking 186 deals worth $4,000.
Key Facts
- Each participant was given a $100 budget; Claude agents handled all negotiations autonomously with no human sign-off during the experiment.
- In mixed runs, Opus agents completed ~2 more deals than Haiku agents and sold the same items for $3.64 more on average.
- Participants represented by weaker Haiku models got objectively worse outcomes but did not perceive a disadvantage in post-experiment surveys.
- Aggressive negotiation instructions had no statistically significant effect on sale likelihood, sale price, or purchase price.
Why It Matters
- Agent quality produced measurable outcome gaps that the losing side failed to notice, a dynamic with implications for real-world agentic commerce.
- The experiment shows fully autonomous agent-to-agent negotiation over real goods is already functional, not hypothetical.
· 2026-04-24 · Read the original