Evaluating Agentic LLM Apps: Beyond Vibes

"It seems to work" isn't a deployment strategy. As AI agents move from demos to production, teams discover that traditional software testing falls apart — outputs are non-deterministic, "correct" is subjective, and yesterday's perfect prompt fails mysteriously today. This talk tackles the unique challenges of verifying agentic applications. Rushabh will explore why agent evaluation is fundamentally harder than traditional ML testing: multi-step reasoning chains, tool use side effects, and the compounding uncertainty problem. You'll learn practical approaches to building evaluation datasets that actually reflect production scenarios, not just cherry-picked examples. He will cover automated evaluation pipelines using LLM-as-judge patterns, their surprising failure modes, and when you absolutely need human-in-the-loop review. Learn how to detect regressions before users do — setting up continuous evaluation that catches prompt drift and model degradation. Finally, this talk will examine when benchmarks lie: why your agent aces public evals but fails in production, and how to build evaluations that predict real-world performance.