26 Agents, 720 Simulated Hours

Sentiment analysis tells you what people said. It does not tell you how a narrative forms, who amplifies it, or whether two groups converge or polarise over thirty days. That is a question about interaction, and you cannot get it from a corpus of static posts — you have to simulate the population and let it run.

A seed document of 1,450 words drawn from public reporting is used to extract an ontology, which becomes a GraphRAG knowledge graph of 121 entities and 158 relationships. Agent personas are synthesised against that graph — laid-off employees, surviving workers, investors, journalists, labour advocates — then dropped into a dual-platform environment modelling Twitter/X and LinkedIn simultaneously, since the same person performs differently on each. The simulation runs 720 hours across 40 rounds, with GraphRAG serving as agent memory so personas stay coherent.

This simulates a hypothetical scenario, so there is no ground truth to validate against — the narratives it surfaces are plausible and internally consistent, not verified predictions. Agent populations built from LLMs also inherit the model's priors about how each group talks, which risks producing a well-articulated stereotype rather than a finding. I would treat the output as hypothesis generation, never evidence.

Three dominant narratives emerged: anger framing layoffs as trading human welfare for AI infrastructure; survivor guilt showing public silence paired with private anxiety; and investor-worker polarisation that intensified over the thirty days rather than converging — the last being the one genuinely counter-intuitive result, since narratives more often decay toward consensus.