Small populations lie politely. Large ones tell the truth. A field report from pushing the simulator past comfortable limits.
Aiman Showkat Khan
Founder · April 11, 2026At 50 agents, the simulator agrees with your intuition. At 500, it starts to surprise you. At 10,000, it shows you the thing you were not looking for — the small, stubborn cluster that refuses the narrative and eventually drags the median with it.
Everything. Token budgets, orchestration, memory persistence, and a lot of engineering judgment about when an agent actually needs to "think" and when it can coast on priors. We ended up inventing a tiered thinking model so that 10,000 agents cost less than 500 did a year ago.
“Once you have watched 10,000 agents disagree about a fuel tax for an hour, a briefing deck never looks quite the same again.”