Taming Rogue AI Agents with Observability-Driven Evaluation

Learn to control unpredictable AI agent behavior through a systematic observability-driven evaluation framework in this 16-minute conference talk. Discover how LLM agents can drift into failure modes when prompts, retrieval systems, external data sources, and policies interact in unexpected ways. Explore a repeatable, metric-driven methodology for detecting problematic behaviors in production agentic systems, diagnosing the root causes of these issues, and implementing effective corrections at scale. Gain practical insights into monitoring and evaluating LLMs and AI agents to prevent costly failures and maintain reliable performance in real-world deployments.