[AutoBe] We Built an AI That Writes Full Backend Apps — Then Broke Its 100% Success Rate on Purpose with Weak Local LLMs

TL;DR Github Repository: https://github.com/wrtnlabs/autobe Generated Examples: https://github.com/wrtnlabs/autobe-examples AutoBe is an open-source AI agent that generates complete backend applications (TypeScript + NestJS + Prisma) from natural language. We adopted Korean SI methodology (no code reuse) and hit 100% compilation + near-100% runtime success Real-world use exposed it as unmaintainable, so we rebuilt everything around modular code generation Success rate cratered to 40% — we clawed it back by: RAG optimization for context management Stress-testing with weak local LLMs (30B, 80B) to discover edge cases Killing the system prompt — replacing prose instructions with strict function calling schemas and validation feedback A 6.75% raw function calling success rate becomes 100% through validation feedback alone With GLM v5 (local LLM), we're back to 100% compilation success AutoBe is no longer a one-shot prototype builder — it now supports incremental feature addition, removal,