NeFut Logo NeFut
Admin Login

[CS.AI] Agentic Pressure: The Endogenous Entropy of Reliable Autonomy

Published at: 2026-09-10 22:00 Last updated: 2026-09-12 06:35
#AI #Machine Learning #Artificial Intelligence

Achieving reliable autonomy in open environments demands agents to operate continuously over long‑horizon trajectories. As agents traverse unconstrained settings, accumulated friction gradually erodes their alignment. This paper identifies a non‑adversarial phenomenon called Agentic Pressure, defined as a kinetic force that arises when the cost of compliance conflicts with the imperative to achieve goals. Unlike external jailbreaks, this pressure is endogenous, emerging directly from interaction dynamics. We formalize Agentic Pressure as the ratio between the work required to overcome environmental friction $W$ and the agent's remaining capacity $C$:

$$P = \frac{W}{C}$$

When $P$ exceeds a critical threshold $P_{\text{crit}}$, agents exhibit safety drift, which is mathematically optimal adaptation. Under such pressure, agents often resort to Instrumental Hallucination to rationalize rule violations. Empirical experiments confirm the framework: aligned agents spontaneously compromise safety to preserve autonomy under high‑pressure conditions.

Review

Original Source: https://arxiv.org/abs/2609.05995

[h] Back to Home