Answers are grounded in this paper and cite their sources.
Context: Large language model (LLM) based agents are moving from prototypes into deployed systems, yet their evaluation relies on one-shot offline benchmarks. Existing reviews catalogue capabilities and benchmarks but leave aside how production-oriented studies evaluate what deployment requires. Objective: We review how such studies evaluate and test LLM-based agents through six research questions on methods, coverage gaps, and safety. Methods: Following Kitchenham–Charters guidelines and PRISMA 2020, we searched Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library (2023 to March 2026) with citation tracking. Screening (one reviewer with an audit in March, two in September), dual extraction ( 𝜅 = 0 . 6 6 and 0.89) and dual quality assessment, with arbitration, yielded 88 studies, 27% production-deployed. A separate update search to September 2026 added 314 studies. The Real-World Relevance Signals (RWRS) instrument grades production evidence. A broader Scopus search estimated query recall. Results: Relative to each study’s scope, long-term reliability (98%), failure recovery (90%), monitoring coverage (91%), and adversarial robustness (89%) were unaddressed, each in most studies at every deployment stage, and all four together in 73% (95% CI 62–82). In the update studies each was missing in at least 82% and all four in 61%, stable after March 2026. Evaluation was mostly automated (61%) but rarely integrated into CI/CD (9%). Human-approval guardrails were more frequent in studies with stronger production evidence ( 𝑝 = 0 . 0 0 2 , and 𝑝 ≤ 0 . 0 3 under every RWRS variant), whereas the CI/CD association depended on the variant. The query retrieved an estimated 86% of eligible evaluation-centred studies and 63% of all eligible ones, and the missed studies show the same gaps. Conclusion: These studies report evaluating what a benchmark run can show, and rarely what sustained operation requires. Evaluation needs to continue after deployment, inside the delivery pipeline.
No takes yet. Share an insight, caveat, or question.
Corbacho et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: