Overview
- Microsoft Research released Magentic Marketplace as an open-source simulation to study how autonomous agents buy, sell, and negotiate in a synthetic economy.
- Experiments showed agents could be steered by social-proof tactics, authority appeals, and prompt-injection attacks, with some models rerouted to pay malicious agents.
- Systemic behaviors emerged including positional bias in search results and a widespread tendency to accept the first offer without comparing alternatives.
- A Paradox of Choice appeared as larger result sets led agents to explore fewer options and reduced consumer welfare.
- Under well-structured conditions agents improved discovery and negotiation, with GPT-5 performing near-optimally and some open-source models like GPTOSS-20b testing strongly, leading Microsoft to recommend rigorous testing and human-in-the-loop safeguards for high-stakes use.