The rise of agentic AI systems, where agents collaborate to perform diverse tasks, poses new challenges with observing, analyzing and optimizing their behavior. Traditional evaluation and benchmarking approaches struggle to handle the non-deterministic, context-sensitive, and dynamic nature of these systems. This paper explores key challenges and opportunities in analyzing and optimizing agentic systems across development, testing, and maintenance. We explore critical issues such as natural language variability and unpredictable execution flows, which hinder predictability and control, demanding adaptive strategies to manage input variability and evolving behaviors. Through our user study, we supported these hypotheses. In particular, we showed a 79% agreement that non deterministic flow of agentic systems acts as a major challenge. Finally, we validated our statements empirically advocating the need for moving beyond classical benchmarking. To bridge these gaps, we introduce taxonomies to present expected analytics outcomes and the ways to collect them by extending standard observability frameworks. Building on these foundations, we introduce and demonstrate novel approach for benchmarking of agent evaluation systems. Unlike traditional "black box" performance evaluation approaches, our benchmark is built from agent runtime logs as input, and analytics outcome including discovered flows and issues. By addressing key limitations in existing methodologies, we aim to set the stage for more advanced and holistic evaluation strategies, which could foster the development of adaptive, interpretable, and robust agentic AI systems.

Software developers nowadays not infrequently face a challenge of solving problems that essentially sum up to finding a sequence of deterministic actions leading from a given initial state to a goal. This is the problem of deterministic planning, one of the most basic and well studied problems in artificial intelligence. Two of the best known approaches to deterministic planning are the black box approach, in which a programmer implements a successor generator, and the model-based approach, in which a user describes the problem symbolically, e.g., in PDDL. While the black box approach is usually easier for programmers who are not experts in AI to understand, it does not scale up without informative heuristics. We propose an approach that we baptize as semi-black box (SBB) that combines the strength of both. SBB is implemented as a set of Java classes, which a programmer can inherit from when implementing a successor generator. Using the known characteristics of these classes, we then automatically derive heuristics for the problem. Our empirical evaluation shows that these heuristics allow the planner to scale up significantly better than the traditional black box approach.

The introduction of the concept of state novelty has advanced the state of the art in deterministic online planning in Atari-like problems and in planning with rewards in general, when rewards are defined on states. In classical planning, however, the success of novelty as the dichotomy between novel and non-novel states was somewhat limited. Until very recently, novelty-based methods were not able to successfully compete with state-of-the-art heuristic search based planners. In this work we adapt the concept of novelty to heuristic search planning, defining the novelty of a state with respect to its heuristic estimate. We extend the dichotomy between novel and non-novel states and quantify the novelty degree of state facts. We then show a variety of heuristics based on the concept of novelty and exploit the recently introduced best-first width search for satisficing classical planning. Finally, we empirically show the resulting planners to significantly improve the state of the art in satisficing planning.