Role-Based Model Allocation Each agent role in the simulation is assigned to a specific LLM tier based on the complexity and importance of its function.
Role
Model
Tier
Purpose
Genesis Architect GPT-4o / Claude Opus
Premium
World-building, persona generation, scenario design
Main Debater GPT-4o
Premium
Core arguments, logical reasoning, position defense
Reaction Agent GPT-4o-mini
Standard
Quick responses, emotional reactions, crowd behavior
Chaos Joker Claude Sonnet
Mid
Contrarian positions, chaos injection, devil's advocate
Searcher Gemini Flash + Tavily
Standard
External data retrieval, fact-checking, news injection
Librarian GPT-4o-mini
Standard
Wiki maintenance, summarization, knowledge graph updates
Cost Optimization Strategy
Premium tier (GPT-4o, Claude Opus): Used only for genesis events, key debates, and critical decision pointsStandard tier (GPT-4o-mini, Gemini Flash): Used for bulk conversations, reactions, and maintenance tasksTarget ratio: 20% premium / 80% standard calls
Wiki updates are batched (every N conversation turns, not real-time)
Knowledge graph reconstruction runs on schedule, not per-interaction
External data injection is rate-limited (configurable interval)
Agent persona definitions cached in Redis
Frequently referenced wiki pages cached with TTL
Conversation context windows managed with sliding window + summary
Scenario
Agents
Duration
Est. Cost
Small (10 agents, 1 epoch)
10
~1 hour
$2–5
Medium (50 agents, 5 epochs)
50
~6 hours
$15–30
Large (200 agents, 20 epochs)
200
~24 hours
$80–150