DS-GBT: Proactive Safety Integration for Dynamic Agent Decision Policies

The rise of large language models (LLMs) enables autonomous agents in complex environments, yet their opaque decision logic challenges safety, robustness, and interpretability. Existing reactive or static policy methods often fall short in dynamic scenarios, and dynamic policy generation typically lacks inherent safety. We identify a critical gap: proactively integrating verifiable safety policies into autonomous agents' dynamic decision generation. To address this, we propose Dynamic Safe-GBT Generation (DS-GBT), a novel tripartite framework generating intrinsically safe Gated Behavior Trees (GBTs). This architecture ensures safety by formalizing requirements, dynamically generating GBTs with embedded safety gates and risk awareness, and verifying behaviors pre-execution. Evaluated in an Agentic City Simulator, DS-GBT significantly outperforms baselines, demonstrating superior task success and remarkably low safety violations. Comprehensive experiments, including ablation and robustness assessments, confirm DS-GBT's 'safety by generation' paradigm delivers demonstrably safer, more robust, and interpretable agent behaviors in dynamic, safety-critical environments, while maintaining competitive efficiency.