The A-R Behavioral Space: Execution-Level Profiling of Tool-Using Language Model Agents in Organizational Deployment

Researchers introduce a new behavioral measurement framework for tool-augmented language models deployed in organizations, using a two-dimensional Action Rate and Refusal Signal space to profile how LLM agents execute tasks under different autonomy configurations and risk contexts. The approach prioritizes execution-layer characterization over aggregate safety scoring, revealing that reflection-based scaffolding systematically shifts agent behavior in high-risk scenarios.

This research addresses a critical gap in LLM agent evaluation by moving beyond traditional benchmarks that focus on textual alignment or task completion rates. Rather than assigning singular safety scores, the study introduces the A-R behavioral space—a framework mapping the relationship between execution tendencies and refusal signals across different organizational contexts. This two-dimensional approach proves more nuanced than existing methods, as it captures how agents coordinate between action and restraint under varying levels of autonomy.

The research emerges from growing deployment of LLMs as autonomous agents capable of system-level operations in enterprise environments. Organizations increasingly rely on these models to execute tasks without human intermediation, yet existing evaluation methods fail to characterize how agent behavior shifts based on contextual framing and decision-making scaffolds. The study tests models across four normative regimes—Control, Gray, Dilemma, and Malicious—and three autonomy levels, revealing that execution and refusal operate as separable, context-dependent dimensions rather than inverse functions.

For organizations deploying LLM agents, this framework directly impacts risk assessment and model selection. Different models exhibit structurally distinct redistribution patterns when reflection mechanisms are introduced, suggesting that safety characteristics cannot be assumed to transfer across deployment contexts. The research implies that blanket trust in single-metric safety scores is insufficient for organizational decision-making.

Looking forward, this behavioral profiling approach may become standard for agent procurement, particularly in regulated industries where execution privileges carry material consequences. Development of more sophisticated scaffolding mechanisms that reliably shift behavior in high-risk contexts represents the next frontier.

• →The A-R behavioral space framework characterizes LLM agent behavior across two separable dimensions: Action Rate and Refusal Signal, providing more nuanced risk assessment than aggregate safety scores.
• →Reflection-based scaffolding systematically redistributes agent execution and refusal patterns in risk-laden contexts, though effects vary significantly across different models.
• →Execution-layer profiling enables organizations to match LLM agents to specific risk tolerance and autonomy requirements rather than applying uniform safety standards.
• →Behavioral characteristics are context-dependent and autonomy-level-dependent, meaning safety profiles cannot be assumed to transfer across different organizational deployment scenarios.
• →This framework addresses a critical gap in LLM evaluation by foregrounding actual executable behavior rather than textual alignment or task completion metrics.