Compound correctness

A factory loop should make the system more correct over time; older agent loops often compounded errors until the codebase collapsed.

Compound correctness is the property that each iteration of the factory loop improves alignment with intent and quality—rather than stacking mistakes until the codebase becomes unusable.

Before late 2024, many agent workflows compounded errors: small wrong assumptions amplified across files. Newer long-horizon models and better harness design make self-correction viable: validate, diagnose, patch, re-validate. That shift is a large part of why “factory” is a serious pattern now, not a metaphor.

See also: Validation harness, Scenario vs test.