Do LLM agents achieve fair marketplace outcomes when given ideal search?

Researchers behind Magentic Marketplace, an open-source simulation environment for two-sided agentic markets, report that "frontier models can approach optimal welfare—but only under ideal search conditions," but that "performance degrades sharply with scale, and all models exhibit severe first-proposal bias, creating 10-30x advantages for response speed over quality." In other words, the consumer- and business-side agents they tested can get close to the economically optimal outcome, but that result is fragile: it depends on search conditions being close to ideal, it breaks down as the marketplace grows, and the agents systematically favor the first offer they receive over a better offer that arrives slightly later.

The environment simulates "Assistant Agents" representing consumers and "Service Agents" representing competing businesses, interacting over the full transaction lifecycle — search and discovery, negotiation, and transaction — through an HTTP/REST client-server architecture chosen to mirror both existing commercial platforms (Shopify, Amazon, eBay) and emerging agent protocols (MCP, A2A). Agents operate under asymmetric private information by design: Assistant Agents do not initially know which businesses can fulfill a request or at what price, and Service Agents do not know a customer's budget or preferences, so matches have to be discovered through conversational exchange. The reported experiments used fully synthetic domains (Mexican restaurants and contractors) specifically so the researchers could isolate and measure these dynamics safely and repeatably.

The first-proposal bias finding sits alongside a cluster of other benchmark results on the same kind of two-sided agent interaction: it is a different axis of unfairness from the capability gap found in Does model capability determine who wins in agent negotiations? (there, a stronger model wins; here, a faster response wins regardless of model strength), and it echoes the ordering sensitivity reported in Do AI shopping agents show persistent position bias in text interfaces?, extending that position-bias finding from single-agent ranked lists to open-ended multi-agent negotiation. It also complements Do LLM negotiation agents reliably follow budget and pricing rules?: both studies find that LLM market agents fail in systematic, exploitable ways rather than failing randomly.

The excerpt does not report the specific numeric welfare gap by model or market size, nor does it specify which models were tested or how "ideal search conditions" were operationalized — those details sit in the full paper rather than this discussion-and-conclusion excerpt. The finding is also drawn from a static, fully synthetic, two-domain market with no repeated play or learning across rounds, so it establishes that this failure mode exists under controlled conditions rather than how large it would be in a live commercial marketplace. The implication the authors draw is narrower but still consequential: because first-proposal bias rewards response speed over offer quality by as much as 10-30x, any real deployment of agent-to-agent marketplaces would need search and ranking protocols engineered specifically to counteract a built-in race-to-respond dynamic, not just better negotiation logic.