The increasing sophistication of Artificial Intelligence (AI) systems necessitates a rigorous, multi-dimensional evaluation paradigm that surpasses conventional automated metrics and subjective human assessments. This paper introduces Jo.E (Joint Evaluation), a structured evaluation framework that integrates human expertise, AI agents, and Large Language Models (LLMs) to systematically assess AI systems across critical dimensions: accuracy, robustness, fairness, and ethical compliance. Building on methodologies such as "Agent-as-a-Judge" and "LLM-as-a-Judge", Jo.E provides a principled approach to identifying and mitigating AI risks through a tiered evaluation process. We validate this framework through controlled experiments on commercial models (GPT-40, Llama 3.2, and Phi 3), demonstrating its capacity to detect model vulnerabilities that single-method evaluations miss. The framework’s key innovation lies in its structured information flow between evaluation tiers, enabling targeted human expert involvement where automated methods are insufficient. This creates a scalable, reproducible evaluation methodology with comprehensive coverage of critical AI safety dimensions. Our experimental results show that Jo.E successfully identified 22% more adversarial vulnerabilities and 18% more ethical concerns than standalone evaluation approaches while reducing human expert time requirements by 54%.
Agentic systems have transformed how Large Language Models (LLMs) can be leveraged to create autonomous systems with goal-directed behaviors, consisting of multi-step planning and the ability to interact with different environments. These systems differ fundamentally from traditional machine learning models, both in architecture and deployment, introducing unique AI safety challenges, including goal misalignment, compounding decision errors, and coordination risks among interacting agents, that necessitate embedding interpretability and explainability by design to ensure traceability and accountability across their autonomous behaviors. Current interpretability techniques, developed primarily for static models, show limitations when applied to agentic systems. The temporal dynamics, compounding decisions, and context-dependent behaviors of agentic systems demand new analytical approaches. This paper assesses the suitability and limitations of existing interpretability methods in the context of agentic systems, identifying gaps in their capacity to provide meaningful insight into agent decision-making. We propose future directions for developing interpretability techniques specifically designed for agentic systems, pinpointing where interpretability is required to embed oversight mechanisms across the agent lifecycle from goal formation, through environmental interaction, to outcome evaluation. These advances are essential to ensure the safe and accountable deployment of agentic AI systems.