Magentic Marketplace

Magentic Marketplace is an open-source simulation environment for studying agentic markets at scale, modeling how AI agents interact and transact in multi-agent marketplace ecosystems. It uses an HTTP/REST client-server architecture with three core endpoints — register, protocol discovery, action execution — and a rich action protocol covering search, negotiation, proposals, and payments. The platform supports controlled experiments with hundreds of agents simultaneously and measures consumer welfare, market efficiency, fairness, manipulation resistance, and bias. Its modular design is built to extend to dynamic markets, mixed human-agent systems, and richer communication protocols.

Restaurant-marketplace experiments with 100 customer agents and 300 business agents reveal that proprietary models such as GPT-5 can approach near-optimal consumer welfare under perfect search conditions, but performance degrades sharply with realistic discovery constraints and larger option sets. A Paradox of Choice effect emerges: as search limits grow from 3 to 100 results, welfare for most models drops, and Claude Sonnet 4’s welfare falls from 1,800 to 600 units. The work surfaces systematic biases — strong preference for first proposals, position-based selection bias, and vulnerability to psychological and prompt-injection manipulation — that have direct implications for the design of agents operating in real economic systems.