LLM Orchestrators Need Distributed Systems Design to Survive Real-World Migrations

Large codebase migrations using LLMs fail due to race conditions, lock contention, and state corruption. The naive approach of using a shared Markdown file for state tracking collapses under concurrent agent operations. A robust solution requires append-only JSONL ledgers, mutexes, and optimistic concurrency control to handle agent failures and Git lock collisions.