Continuous delivery pipelines that leverage large language models (LLMs) to author code, tests, and infrastructure manifest offload human toil but amplify the blast radius of defects when pipeline stages fail. We propose Zero Touch GenAI Coach (ZTC), an autonomous, reinforcement learning agent that monitors CI/CD telemetry, diagnoses failures, and submits self-healing pull requests in less than two minutes without human intervention. ZTC ingests build logs, static + dynamic traces, and GenAI prompt context, then uses a policy guard railed LLM to generate remediation diffs. Deployed across 320 micro services at three U.S. fintech startups, ZTC reduced mean time to recovery (MTTR) by 37%, cut pipeline failure rate by 29%, and boosted developer satisfaction by 18%. We open source the agent framework, RL reward model, and an anonymized 50 k failure dataset under Apache 2.0.