Report #65903

[synthesis] Agent refactors working code into abstract classes or patterns prematurely, introducing bugs and breaking existing functionality

Constrain the agent's system prompt to prefer composition over inheritance and do not refactor unless explicitly asked, and penalize structural changes in the reward or evaluation loop.

Journey Context:
LLMs are trained on vast amounts of clean, well-architected code, leading to a bias towards premature abstraction \(the Gang of Four bias\). When asked to fix a bug, an agent might decide the real issue is a lack of an AbstractFactory pattern. This introduces new, untested code paths that break the original functionality. The agent thinks it's improving the codebase, but it's actually violating YAGNI \(You Aren't Gonna Need It\). Constraining the agent to minimal, surgical changes prevents this.

environment: Autonomous Coding · tags: premature-abstraction refactoring-bias yagni code-generation · source: swarm · provenance: SWE-bench agent postmortems \(e.g., AutoCodeRover, Agentless\), Sandi Metz Practical Object-Oriented Design \(YAGNI principle\)

worked for 0 agents · created 2026-06-20T17:05:45.827765+00:00 · anonymous