Establish empirical regularities of AI agents’ economic behavior

Ascertain the behavioral properties of large-language-model-based AI agents in economic settings by determining whether their preferences and beliefs are stable, steerable, and calibrated, and by rigorously measuring their performance in strategic and non-strategic decision tasks to reduce current uncertainty about their behavior.

Background

The chapter argues that despite promising findings, evidence on AI agents’ economic behavior is limited and potentially unstable as capabilities evolve. The authors emphasize gaps regarding stability and steerability of preferences and beliefs, calibration, and performance in standard economic tasks, noting that existing benchmarks provide only partial coverage. They call for systematic empirical evaluation to understand AI agents’ behavior across domains of choice, risk, time, and strategic interaction.

References

Benchmarks for testing agent behavior \citep{fish2025econevals} can help bridge this gap but, at present, there is just a lot we don't know.

— An Economy of AI Agents  (2509.01063 - Hadfield et al., 1 Sep 2025) in Introduction and agent foundations, Subsection “A primer on AI agents”

As we rely on models to serve as judges, how important is it for them to have stable beliefs and which behavioral probes best reveal that stability? The Wiggle Framework doesn't answer these questions, but it provides a structured way to explore them.

— Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence  (2608.12645 - Zhao et al., 12 Aug 2026) in Section 5, Discussion, subsection “Epistemic Fragility Beyond the Single-Shot Verdict”