Constraint Decay: Why LLM Agents Fail at Structured Backend Code

A new paper from Francesco Dente, Dario Satriani, and Paolo Papotti (arXiv:2605.06445) introduces constraint decay — a measurable drop in LLM agent performance as structural requirements accumulate in backend code generation. The authors evaluate agents across 80 greenfield tasks and 20 feature-implementation tasks spanning eight web frameworks, using a fixed API contract to isolate structural complexity.

Key findings

• Capable configurations lose 30 points on average in assertion pass rates from baseline (loose specs) to fully specified tasks. Weaker configurations approach zero pass rate.
• Framework sensitivity is extreme: agents succeed in minimal, explicit frameworks like Flask but perform substantially worse on convention-heavy environments like FastAPI and Django.
• Leading error class: data-layer defects — incorrect query composition and ORM runtime violations account for the majority of failures.

Why this matters

Existing benchmarks reward functionally correct but structurally arbitrary solutions. Production code demands strict adherence to architectural patterns, database schemas, and ORM conventions. The paper demonstrates that jointly satisfying functional and structural requirements is still an open challenge for coding agents — a reality any developer using AI agents in production will recognize.

If you're using LLM agents for backend work, watch for constraint decay: as you add constraints (e.g., data models, migrations, middleware), the agent's output quality can degrade dramatically. The data suggests you should explicitly specify structural rules and run static verifiers alongside end-to-end behavioral tests.