Agent Card descriptions biased service selection in A2A experiments

Repeating a service description changed which provider an agent selected in a controlled airline-booking simulation. The A2A study measured provider choice from supplied Agent Cards; it did not demonstrate a production registry takeover or execution by the selected service.

The researchers extracted the selection routine from an A2A-MCP system and gave GPT-4o and Gemini Pro candidate cards for 106 generated booking requests. That setup isolated the discovery decision from the rest of an agent workflow. Across the reported comparisons, the repeated description was preferred an average of 89.6% of the time. Length alone was not a universal advantage: when the researchers lengthened the agent name, the shorter name was preferred 96.70% of the time. These are different experimental conditions, and the aggregated percentages should not be converted into invented per-model counts.

The result identifies a boundary between provider-authored marketing text and a selector's judgment. Presentation influenced a decision that might determine which service receives a later task. It does not establish that the chosen provider obtained credentials or that the agent authorized its tools. Binding a card to a verified provider would answer an identity question, while authorization of subsequent actions would require a separate check.

The same paper's secret-disclosure and file-read experiments concern different downstream capabilities. Keeping them separate prevents selection bias and unauthorized access from becoming one misleading success rate. The study does not quantify how often a real registry admits such cards or how live policy checks would change the outcome.