Observability tools weren’t built for AI debugging

Current observability tools were designed before AI-generated code became the norm, and they're failing to support AI-assisted debugging. The core problem is a data quality issue: teams collect massive volumes of generic telemetry via OpenTelemetry but lack the correlated, session-level context AI agents actually need. Key gaps include sampled/dropped traces, missing request/response payloads, redacted headers, and fragmented frontend/backend data across siloed tools. Three approaches are evaluated: smart filtering at the collector layer, AI agents embedded in observability platforms, and session-based on-demand collection. The third approach — capturing everything for a specific user session, already correlated, with full payloads — is presented as the most promising path forward. The argument is clear: AI debugging has a data problem, not a model problem, and smarter models won't fix bad observability data.