Minimal Footprint as an Optimization Constraint for Safe LLM Agents

Authors: Aarush Ambar

Country: India

Full-text Research PDF File:   View   |   Download

Abstract: The deployment of large language model (LLM) agents in production environments introduces a class of safety risks distinct from those of passive text generation systems. Autonomous agents capable of executing tool calls, managing files, browsing the web, and spawning sub-agents may acquire resources, permissions, and capabilities that significantly exceed task requirements — creating expanded attack surfaces, irreversible side-effects, and cascading failure risks in multi-agent pipelines. The minimal footprint principle, articulated in recent AI safety literature, prescribes that agents should request only necessary permissions, prefer reversible actions, avoid persistent data storage beyond immediate needs, and minimize side-effects. Despite its intuitive appeal, this principle remains informally defined and has not previously been operationalized as a measurable or trainable objective. This paper addresses that gap. A formal multi-dimensional footprint metric F is proposed as a weighted linear combination of five behavioral dimensions: permission scope, data persistence, irreversible side-effect count, sub-agent spawning, and resource duration. This metric is integrated into reinforcement learning training via a reward shaping framework, producing footprint-constrained agents that minimize unnecessary resource acquisition while preserving task performance. Empirical evaluation on the WebArena and ALFWorld benchmarks demonstrates that footprint-constrained training achieves a 43% reduction in measured footprint with only a 6.8 percentage point degradation in task success rate at the recommended operating point. An inference-time footprint scorer is additionally proposed and evaluated, enabling footprint filtering on pre-trained agents without retraining. This work constitutes the first formal operationalization of minimal footprint for LLM agents and contributes an open-source evaluation framework for agentic safety research.

Keywords: LLM Agents, Agentic Safety, Minimal Footprint, Reward Shaping, Reinforcement Learning, AI Safety, Optimization Constraint

Paper Id: 233102

Published On: 2026-05-07

Published In: Volume 14, Issue 3, May-June 2026

Share this