Function Calling Harness: From 6.75% to 100%

The qwen3-coder-next model's first-attempt success rate when generating complex shopping mall backend API data types is only 6.75%. When dealing with nested structures or complex schemas, an LLM's Structured Output remains unstable, and this can even degrade the model's reasoning ability.

The solution lies not in a better model or a more sophisticated prompt, but in Harness engineering outside the model itself. AutoBe is an AI backend auto-generation agent that produces database schemas, API specifications, and even implementation code from natural language conversation alone.

The core infrastructure of this system is Typia. Typia automatically builds a schema, parser, validator, and feedback generator from a single TypeScript type. This forms a deterministic feedback loop that pinpoints the model's errors precisely.

As a result, AutoBe converts the initial 6.75% success rate into a compile success rate of over 99.8%. This is the achievement of a Self-healing approach that wraps a probabilistic LLM in a deterministic verification loop, making it correct its own errors.