arXiv: TDAD — Test-Driven Agentic Development Reduces AI Coding Agent Regressions via Graph-Based Impact Analysis

arXiv·high signal

TDAD introduces a graph-based impact analysis layer that determines which tests are affected by AI-generated code changes before execution, directly addressing the regression problem where coding agents silently break previously-passing tests. The framework integrates into existing coding agent pipelines as a pre-execution gate, using dependency graph traversal to scope test runs to only impacted surface area. This targets one of the most persistent production complaints about coding agents — confident but silent regressions that reach CI undetected.