Constraint Decay: The Fragility of LLM Agents in Back End Code Generation
4 hours ago
- #Backend Code Generation
- #LLM Agents
- #Software Engineering
- LLM agents perform well in autonomous code generation but struggle with strict structural constraints required for production-grade software.
- Existing benchmarks often miss non-functional requirements, leading to functionally correct but structurally flawed solutions.
- The study introduces 'constraint decay'—a significant drop in agent performance as structural requirements accumulate.
- Agents perform better in minimal frameworks (e.g., Flask) and worse in convention-heavy environments (e.g., FastAPI, Django).
- Data-layer defects, such as incorrect query composition and ORM violations, are identified as major root causes of errors.
- Satisfying both functional and structural constraints remains a key challenge for coding agents in backend code generation.