Vibe Coding
Pattern: Function Calling Reliability Becomes Infrastructure — Harnesses, Not Prompt Engineering, Are How You Get 100% Tool Use
AutoBE's Qwen Meetup presentation crystallizes an emerging pattern: reliable function calling requires compiler-level infrastructure (TypeScript AST schema generation, lenient JSON recovery, structured validation feedback), not better prompts. Their harness takes Qwen3-Coder-Next from 6.75% to 99.8% success. This matches the trajectory of Claude Code's own tool use infrastructure (permission rules, hook chains, structured tool schemas) and Anthropic's documented harness design pattern. The implication: 'just prompt it better' is dead for production agent tool use — you need a parsing and validation layer between the model and your tools.
↳ Follow the thread