Software quality has traditionally depended on explicit specifications, testing, and verification. A new arXiv preprint argues that agentic systems—which outsource parts of their logic to large language models—break this foundation. The paper, 'When Code Becomes an LLM Call: An Empirical Study of Call-Site Contracts in Agentic Systems,' examines how these systems handle the contracts that normally govern function calls.
The authors study 'call-site contracts'—the expectations and guarantees at the point where code invokes an LLM instead of a conventional function. Because LLM outputs are probabilistic and not bound by formal specifications, the usual mechanisms for ensuring correctness may not transfer directly. The abstract notes that agentic systems 'replace parts of their logic' with LLM calls, but the full details of the study are not available in the provided excerpt.
The significance is that as AI-driven components become more common in software, the discipline of software quality will need new approaches. The paper is an empirical look at how contracts are (or are not) maintained at these call sites, suggesting that the field is only beginning to understand the implications.