A conference talk by Phil Hetzel from Braintrust comparing traditional observability (uptime, latency, error rates) with agent observability for AI systems. Key differences covered: agents are non-deterministic unlike traditional apps, agent traces are semi-structured, voluminous (up to gigabytes), and contain unstructured text requiring full-text indexing. Agent o11y also requires qualitative metrics like groundedness, tool usage correctness, and brand alignment. The talk highlights that agent observability involves non-technical domain experts (clinicians, lawyers) reviewing traces, unlike traditional o11y. Braintrust built a custom database (moving away from ClickHouse) to handle these workloads, incorporating write-ahead logs, analytical indexes, and Tantivy (a Rust-based Lucene-like full-text index). The future direction includes LLM-powered topic modeling and sentiment analysis over traces to shorten the production-to-fix iteration loop.