Most agentic pipelines are tested on controlled datasets. This session examines one deployed in production across a national wireless carrier, running continuously without manual initiation across hundreds of thousands of geographic contexts. Building a geospatial intelligence system at this scale exposed the realities of production ML: live network telemetry with no clean staging layer, spatial models that had to generate and validate autonomously, and pipeline failures that created real operational blind spots. The talk breaks down the infrastructure decisions that made this viable. Council-style multi-agent validation replaced single-model outputs, catching boundary errors before they propagated. Self-healing logic allowed the system to detect failures, re-query data, and correct itself without stopping the pipeline. Early production revealed failure modes no architecture diagram anticipated. The takeaway isn’t the stack. It’s what changes when your data is live, the system runs continuously, and errors directly impact where field teams are deployed. These constraints redefine validation, self-healing, and how confidence in outputs is communicated. Attendees will leave with a practical framework for agentic pipeline reliability, multi-agent validation, and what “production-ready” actually means at scale.