Structured Feedback Improves Repair in an LLM Agent Loop

VeriHarness implements a code-controlled agent loop that optimizes error correction by providing structured feedback containing failure locations, observed values, and admissible alternatives. In TextWorld game environments, this structured approach increased terminal success for Qwen2.5-Coder-14B from 28% to 72% and for Llama-3.1-8B from 16% to 58%.