Agentic Observability - Making LLM Apps Debuggable, Trustworthy, and Scalable by Krishna Gade

In this monologue podcast, Krishna Gade, the founder and CEO of Fiddler AI, introduces the concept of agentic observability and its importance in ensuring the reliability and trustworthiness of AI agentic applications. He discusses the shift from traditional, deterministic software to model-based, non-deterministic agentic AI, highlighting the complexities of monitoring multi-agent workflows compared to single-agent systems. Gade argues for a new approach to application performance management (APM 2.0) that combines traditional APM metrics with model observability metrics to provide hierarchical visibility into agentic workflows. He emphasizes the need to internalize observability within agentic workflows, focusing on reflection, tool usage, and agent collaboration, and introduces new metrics to track these aspects. Gade concludes by advocating for the adoption of agentic observability to enhance the reliability, debuggability, and trustworthiness of AI applications, enabling better alignment with business policies and facilitating live intervention for course correction.