Most observability platforms were designed for infrastructure metrics — CPU utilization, stack traces — data with a short useful lifespan. AI data is fundamentally different. A conversation from six months ago might reveal a failure pattern you only recognize today. The reasoning traces from your best-performing agent sessions are training signals for the next iteration of agents or employees. Sending this to a 30-day retention window is, as Harvey puts it, “like writing your institutional knowledge on a whiteboard and erasing it every month.”
This challenges the default assumption behind Observability is the missing discipline for agent systems — you can't improve what you can't measure: it’s not enough to measure — you must also retain. The architectural implication is that AI traces should live in your own data lake in open formats (Parquet, Iceberg), not in a provider’s proprietary silo. This connects directly to Decision traces are the missing data layer — a trillion-dollar gap — the traces Dhinakaran describes ARE the missing data layer, and their value compounds over time rather than decaying.