Validating Group-Conditioned Behavior in LLM Agent Networks: Separating Group History from Cues and Reputation

Open agent networks create settings in which agents interact repeatedly, remember past partners, and encounter others through shared categories such as seller networks, marketplaces, brand ecosystems, fulfillment channels, or user communities. Over time, these interactions may shape how agents treat not only known individuals but also other members of the same group. Measuring this kind of group-conditioned behavior is challenging because several mechanisms can produce similar observations, including direct responses to group labels, individual reputation, reciprocity, and information retained from prior interactions. We propose a construct-validity framework for distinguishing these mechanisms and measuring when group history has an effect beyond the current cue. The framework compares a history-carrying agent with a matched copy that faces the same current decision without the relevant interaction history, and combines this comparison with hidden-label, memory-scrubbing, novel-member, and reassignment probes. In a pilot with GPT-4.1-mini, visible group labels have a strong immediate effect, while the additional effect of accumulated history is smaller. Much of the behavior that persists after labels are hidden can be accounted for by individual reciprocity, while contested group interaction also changes how agents treat previously unseen members of a category. These results illustrate why separating individual reputation from category-level generalization matters for agentic networks, where past experience with one set of agents may influence how new members of the same group are treated.