Aditya Mulik shares how Walmart scaled resilient multi-agent LLM batch workflows across 4,700 stores, cutting monthly inference costs by 90% while enforcing semantic validation and bounded retries.