Jorge Díaz-Villegas

Agentic Software Engineering (ASE) systems depend on decisions about context, reasoning, coordination, tool use, execution, and verification. This systematic mapping study maps problems, design decisions, operationalizations, and evidence in peer-reviewed studies of LLM- or foundation-model-based agentic systems in Software Engineering from 1 January 2024 to 13 July 2026. ACM Digital Library, IEEE Xplore, Scopus, and Springer Nature Link were searched on 13 July 2026 using the same Boolean string. From 1069 records, 94 studies were included. A frozen, single-paper-isolated extraction protocol identified 1018 design decisions, followed by deterministic multi-label synthesis; no formal study-level risk-of-bias assessment was performed. Of these decisions, 633 were linked to 656 problems or objectives, and 913 (89.7%) had empirical or comparative evaluation records. Only 19.8% of evaluation records attributed evidence to an individual decision, whereas 52.3% assessed complete systems; multi-agent configurations represented 68.7% of problem-linked decisions. Heterogeneous metrics and system-level evaluation limit inference about effectiveness. The findings expose a gap between evaluation coverage and decision-level attribution and motivate ablations and human-centered evaluations. The review received no external funding and was not registered.

Agentic Software Engineering systems depend not only on model choice but also on decisions about context, reasoning, coordination, tool use, execution, and verification. The literature provides a fragmented account of how motivating problems connect to design decisions, operationalizations, and evidence. We address this gap through a systematic mapping study of 94 peer-reviewed primary studies selected from 1,069 records. Evidence-grounded extraction under a frozen, single-paper-isolated protocol identified 1,018 design decisions, of which 633 were explicitly linked to 656 problem or objective mentions. One author audited a random sample of 10 studies comprising 106 decisions and 637 evidence records and found no candidate- or evidence-level discrepancies. Deterministic multi-label codebooks were constructed from the extracted vocabulary and applied using complete token-boundary matching, preserving traceability to source evidence. Although 913 decisions (89.7%) carry an empirical or comparative evaluation record, only 19.8% of evaluation records attribute evidence to an individual decision, whereas 52.3% assess complete systems. Multi-agent configurations account for 68.7% of problem-linked decisions, while hybrid configurations and early life-cycle activities remain sparsely represented. These findings expose a substantial gap between evaluation coverage and decision-level attribution and motivate more targeted ablations and human-centered evaluations.

Large language models and AI agents are extending data-engineering automation beyond isolated artifact generation toward end-to-end processes in which agents interpret requirements, select data, generate transformations, invoke tools, validate results, and communicate analytical outputs. This shift introduces risks that conventional notions of data quality and execution success do not fully capture. A dataset may satisfy established quality standards, and a generated query may execute without technical errors, while the agent still selects an incorrect metric, combines incompatible analytical grains, accesses unauthorized data, or draws conclusions that are insufficiently supported by evidence. This paper develops a conceptual framework for Agentic Data Engineering centered on Quality of Agentic Data Use, defined as the extent to which an agent uses and communicates data in accordance with task, semantic, quality, security, governance, and provenance requirements. An evidence-informed analysis of Data Contracts, Semantic Layers, Data Quality, Guardrails, AI Governance, and Data Provenance shows that these foundations provide essential but fragmented capabilities. The proposed framework integrates and extends them through four core artifacts: Agentic Data Contracts, Agentic Expectations, Agentic Data Provenance, and Agentic Data Governance. It also introduces an execution lifecycle, a reference architecture, a failure taxonomy, and a multidimensional evaluation framework. A governed sales-analysis scenario illustrates how the proposed artifacts interact throughout an agent-mediated data process. In addition, a controlled Databricks prototype and a complementary benchmark comprising 10 cases and 40 executions demonstrate the framework’s technical feasibility and support the independent computation of enforcement indicators. The benchmark highlights the value of separating generation from validation while also showing that the current validation and automated-repair mechanisms require further calibration. These preliminary findings do not establish generalized improvements in safety, correctness, or reliability. Rather, they provide an operational foundation for broader empirical evaluation of trustworthy agent-mediated data-engineering processes.

The integration of large language models and autonomous agents into data engineering is shifting data systems from human-specified pipelines to processes in which agents interpret requirements, select sources, generate transformations, invoke tools, and communicate analytical results. This shift introduces a governance gap: a dataset may satisfy established quality requirements and a generated query may execute successfully while the agent applies an incorrect metric, combines incompatible grains, uses unauthorized sources, or produces conclusions without evidential support. We develop a conceptual framework for Agentic Data Engineering, the engineering domain concerned with specifying, validating, governing, and auditing data processes in which AI agents operate under delegated authority. An evidence-informed analysis of six foundations—Data Contracts, Semantic Layers, Data Quality, Guardrails, AI Governance, and Data Provenance—shows that existing controls are essential but fragmented. The framework integrates them through four artifacts: Agentic Data Contracts, Agentic Expectations, Agentic Data Provenance, and Agentic Data Governance, and defines Quality of Agentic Data Use as the central evaluative construct. A lifecycle model, a reference architecture, a failure taxonomy, a multidimensional evaluation scheme, and an illustrative governed sales-analysis scenario operationalize the framework and provide a foundation for the empirical validation of trustworthy agentic data systems.