Abstract

As agentic systems increasingly incorporate multi-step processes, evaluation remains focused on predictive accuracy rather than process structure. This creates a gap in understanding how decisions are formed, particularly in systems that integrate multiple components. This gap is particularly critical in decision support contexts, where understanding how decisions are formed is as important as the decisions themselves. This work proposes a structural evaluation framework for intelligent agentic systems based on sequence and network analysis. Execution is modeled as state sequences and as dynamic interaction graphs, where nodes represent processes and control states, and edges represent transitions during decision formation. Sequence similarity is analyzed using Optimal Matching (OM) distance to assess stability and consistency of trajectories. Network-based measures, including centrality and motif frequency, are used to characterize coordination patterns, control structures, and progression toward decision commitment. Unlike traditional evaluation approaches, the framework explicitly models and evaluates the structure of execution processes rather than relying solely on outcome-based metrics. The framework is domain-agnostic and can be applied to a broad class of agentic and multi-step decision systems. The study contributes a structural evaluation perspective for intelligent agentic systems, shifting assessment from predictive outcomes to the organization and dynamics of processes. This is particularly relevant for emerging agentic systems that involve iterative planning, tool use, and multi-stage execution.

Download

Share

COinS