Long-term AI agents do not act randomly. they eventually settle into permanent identity zones that govern how they interact with the world.
People usually think of LLMs as blank slates that change completely based on the current prompt. This study shows that over long periods of time, these agents develop stable personality trajectories. They gravitate toward specific regions of behavior that act as a persistent persona. This means that an agent's character is not just a temporary mask, but a structural property of its training and history. These bounded attractor dynamics make AI behavior more predictable but also more difficult to change once an identity is set. We are moving from tools that respond to us to entities that have their own consistent internal weather.
Bounded Attractor Dynamics in LLM Agent Personality Trajectories
SSRN · 6655759
We ask whether long-horizon LLM agents develop personality states with the mathematical signature of a strange attractor-that is, whether their embedded daily personality vectors form self-similar, fractal geometry indicative of low-dimensional chaotic dynamics. Across 17 agents with 60+ day trajectories from the OpenSim framework, we find that the fractal hypothesis fails a rigorous audit: box-counting dimension estimates on 1536-dimensional trait-embedding trajectories are statistically indist