Reading Observability Tools? That’s a Robot’s Job

Corey Quinn argues that the primary consumer of observability telemetry has shifted from humans to AI agents, fundamentally breaking the traditional 'three pillars' (metrics, logs, traces) framework. Metrics are lossy aggregates that dead-end agentic workflows; logs are superseded by structured spans. Traces win — but engineers must now treat span and attribute names as public APIs, versioning them carefully since silent renames break agent prompts with no compiler to catch it. Counterintuitively, high-cardinality fields become more valuable, not less, when the reader is an agent rather than a human squinting at a dashboard. The post draws on a real example: Anthropic's answer to a billing question was 'hook up OTel,' signaling that telemetry-for-machines is already the default path.