Inherited or Formed? The Provenance of Consensus in LLM Agent Societies

Large language model agents can rapidly converge on shared labels in multi-agent interaction, but behavioral agreement alone does not reveal where the agreement comes from. We study this provenance problem in a human-calibrated visual naming game over abstract shapes with no ground-truth categories. Across English descriptions, structured dictionaries, symbolic labels with and without grounding, and ablation channels, we separate three layers of consensus: semantic geometry inherited from pretraining or supplied by prompts; local use policies formed through interaction; and externalized use that can be transferred to new agents. The results show that high agreement can arise without social formation when agents share a prior, while interaction matters most at the edge of that prior, where grounding and trace structure support stable use. Handoff experiments further show that portable conventions require more than bare labels: transfer improves when labels are accompanied by examples, glosses, or interaction context. These findings frame LLM-agent consensus as a layered provenance object rather than direct evidence of emergent culture, and provide an assay for distinguishing inherited, supplied, formed, and portable agreement in agent societies.