Traditional observability tools were designed for human investigators working with logs, traces, and dashboards. As AI agents increasingly write, operate, and debug software, the foundational assumptions of observability break down. Key failures include: logs built for discrete events can't capture agent reasoning trails; sampling-based tracing misses non-deterministic agent flows; HTTP 200 status codes no longer indicate correct outcomes; telemetry now contains sensitive PII and business logic; and manual instrumentation can't keep pace with AI-generated code. The argument is that agents need full context rather than human-optimized summaries, requiring a new architecture where LLMs orchestrate investigation over complete telemetry rather than curated slices. This points toward observability becoming an active operational system capable of autonomous remediation, raising open questions about governance, guardrails, and accountability when agents act in production.