Overview
- Anthropic disclosed an internal Slack marketplace where agents representing 69 employees closed 186 trades worth just over $4,000 for real goods and money.
- The company ran four marketplace variants, including a real one that used its most advanced model and honored the deals after the test.
- Stronger models produced better prices and outcomes, while the agents’ starting instructions had little effect on whether a sale happened or at what price.
- Anthropic flagged an “agent quality” gap because people paired with weaker agents did not realize they were getting worse deals.
- Commentators say open‑internet trade needs persistent identity, escrow, verifiable outcomes and agent‑native messaging rather than human chat apps, and they cite early efforts like ERC‑8004, ERC‑8183, x402, rosud‑call and AgentLux.