Constraint Decay: Vulnerability of LLM Agents in Backend Code Generation

LLM agents show excellent performance under loose specifications, but reveal limitations in production-grade software development that involves strict structural constraints such as architectural patterns, databases, and ORMs.

Researchers discovered a phenomenon called 'Constraint Decay', in which an agent's performance degrades sharply as structural requirements accumulate. According to the experimental results, even agents in capable configurations showed an average 30-point drop in assertion pass rate compared to the baseline task, and some configurations had performance converge to nearly zero.

Performance differences were also significant depending on the characteristics of the framework.

• In explicit and simple frameworks like Flask, relatively high success rates were observed.
• In environments with complex conventions like FastAPI and Django, performance was notably low.

The main cause of errors was analyzed to be data layer defects. Incorrect query construction and ORM runtime violations were identified as the key factors undermining agent performance.