Why Evaluating AI Agents is Hard and What You Can Do About It

Building AI agents has become increasingly accessible, but evaluating their performance remains a major challenge . Unlike traditional machine learning models, AI agents operate through multi-step reasoning, tool usage, and dynamic decision-making , making their outputs difficult to measure using standard metrics. As complexity grows, manual testing becomes insufficient and unreliable .

In this session, we will explore why agent evaluation is fundamentally different from LLM evaluation . You’ll learn about the key quality dimensions for evaluating agents and practical strategies such as deterministic evaluation metrics and LLM-as-a-judge approaches . Drawing from real-world experience with production systems, this session provides a practical framework to evaluate and improve AI agents effectively .

This is an insight-driven and practical session designed to help practitioners build more reliable and measurable AI systems .

Key Takeaways:

• Challenges in Agent Evaluation – why traditional evaluation methods fail for AI agents
• Agent vs LLM Evaluation – understanding key differences in behavior and measurement
• Quality Dimensions – what metrics matter when evaluating agent performance
• Evaluation Techniques – deterministic metrics, LLM-as-a-judge, and hybrid approaches
• Practical Framework – how to evaluate real-world AI agents effectively

Rittika Jindal is a Principal Engineer at Thomson Reuters with 17+ years of experience in cloud, data, and AI/ML systems. She builds production-grade AI agents and GenAI applications, specializing in multi-agent architectures, MCP integrations, and evaluation frameworks.

An active voice in the AI community, she shares practical insights on AI engineering and mentors through Women in Big Data and She loves data. She is passionate about building AI systems that don’t just demo well but work reliably in production.