Multi-Agent Forecasting

Multiple agents help only when they add independent information or structured disagreement. Re-running the same prompt at higher temperature usually produces correlated forecasts, not a useful ensemble.

The Problem

Forecast pipelines often report "agent consensus" from several identical agents. On well-specified macro questions, those agents read the same evidence and return nearly identical probabilities. Averaging correlated forecasts gives false confidence unless the system measures diversity, calibrates probabilities, and stress-tests the consensus with opposing arguments.

The Pattern

WRONG

forecasts = [agent.run(question, temperature=0.7) for _ in range(5)]
p_yes = sum(f.p_yes for f in forecasts) / len(forecasts)
print(f"consensus={p_yes:.2%}")