QuoteBench Reveals Parser Failures in LLM Coding Agents

QuoteBench evaluates how serialization and parsing errors in Bash interfaces mask model performance. Testing across 56 one-shot tasks shows that unescaped parsers reduce success rates by 55.4 to 73.2 percentage points, proving that execution failures often stem from transport rather than generation.