Why Most AI Agents Fail - And How to Build Agents You Can Count On | Ep40

Description

Subscribe to our newsletter: https://dataneighbor.substack.com/ DNP Swag Store: https://dataneighbor.creator-spring.com AI is moving fast, but reliable agents are still rare. In this Data Neighbor Podcast, we sit down with Jigyasa Grover, ML Engineer at Uber, author of Sculpting Data for ML: The first act of Machine Learning, and member of Google’s ML Advisory Board, to unpack why most AI agents fail and what it really takes to build ones you can count on. Jigyasa shares how to design, evaluate, and secure reliable agent systems—from memory management and adversarial testing to using human judgment without slowing down innovation. Connect with the team (tell us YouTube sent you!): - Shane Butler: https://linkedin.openinapp.co/b02fe - Sravya Madipalli: https://linkedin.openinapp.co/9be8c - Hai Guan: https://linkedin.openinapp.co/4qi1r Connect with Jigyasa: https://www.linkedin.com/in/jigyasa-grover/ In this episode, Jigyasa explains how agents evolve beyond simple workflows into autonomous systems, why evals are at the heart of reliable AI, and how developers can prevent silent failures through better design, testing, and observability. You'll learn about: -Why most AI agents fail and how to engineer reliability from day one -Workflow agents vs LLM-based agents -How evals, memory hygiene, and adversarial testing improve reliability -When to use traditional ML instead of LLMs -Designing for human judgment, security, and recovery in agent systems Chapters 00:00 Introduction to Jigyasa Grover and her journey 03:09 Exploring the world of AI agents 05:49 Workflow vs LLM agents 08:17 Why AI agents fail 11:05 Memory management in AI agents 13:48 When traditional ML still wins 16:28 The future of AI agents in consumer apps 26:49 Multi-agent systems and security challenges 33:12 How AI is shaping daily work 40:14 Evaluating AI models and their reliability 49:20 The future of trust and human judgment in AI 52:53 Outro #aipodcast #aiagents #aidevelopment #aiengineering #llm #mlops #datascience #agentdesign #workflowagents #memory #evaluation #productstrategy #aiproductmanagement #autonomousagents #aiethics #aideployments #reliableai #dataneighbor #jigyasagrover #agenticai