The authors introduce Incognita, a Concordia-based framework that separates social interaction from grounded execution to evaluate how generative agents handle knowledge partitioned across role-isolated participants. The system routes messages to specialist entities who mediate operations within a deterministic sub-environment, allowing for the assessment of communication as exploration and action as exploitation.
- Incognita-Retail transforms tau-bench retail into a multi-entity environment while preserving final-state reward semantics.
- Three generative agent models were evaluated on 18 tasks stratified by social breadth across 540 trials.
- Success rates rose from 0% to 8.9% and 17.2%, while premature finalization decreased from 100% to 87% and 58%.
- Stronger models elicited more hidden knowledge, contacted more entities, and attempted more grounded writes, though reliability remained low.
These findings demonstrate that socially distributed task environments expose critical agent behaviors such as knowledge elicitation, source selection, and premature completion beliefs before reliable success is achieved.