Abstract
Agriculture is under pressure from climate variability, labour shortages and the need to use water, fertiliser and pesticide more carefully. Artificial intelligence, and agents and multi-agent systems in particular, have been applied to these problems since the 1990s, but a new wave of systems built on large language models has appeared in the last two years. We use the term agentic artificial intelligence, defined operationally in this paper, for software that perceives part of its environment, decides what to do based on that perception, and then acts or advises, where the sequence of steps is determined by the system itself rather than fixed in advance; this covers both the classical agent tradition and the recent language-model wave, and we distinguish the two throughout. This review asks how far agentic approaches, in both traditions, have travelled into agriculture. We searched Scopus for work published between 2020 and 2026 and screened 4111 records against a protocol based on PRISMA 2020. After removing duplicates, off-topic domains and studies that did not report enough method or result detail, 181 studies remained. We coded each one along six dimensions, agricultural domain, agent organisational pattern, artificial intelligence backbone, data source, deployment setting and level of autonomy, and assessed the consistency of the coding through a blinded manual re-coding of 30 randomly selected records. This quality-control exercise assessed the reproducibility of the record-level classification; it was not intended as a full-text validation of every implemented mechanism. Three findings stand out. First, the field as a whole spans more than three decades, but the large-language-model subset is very young: such studies appear only from 2024, and 46 of the 47 of them were published in 2025 or 2026. Second, the capabilities that define agentic behaviour are unevenly reported. Collaboration is reported by 82 percent of studies, while planning and reasoning each appear in 31 percent, and memory and reflection in 6 and 4 percent; these figures describe what abstracts report rather than confirmed implementations, a distinction we treat carefully throughout. Third, tested evidence is thin. Only 33 studies report a field or real deployment, 68 remain conceptual, and just 11 report evaluation across more than one season or period. We contribute a taxonomy that separates classical and contemporary agentic approaches, a direct comparison between the two, a reported-capability matrix reported with raw counts as well as percentages, and a roadmap for future work that we label clearly as our own synthesis rather than a direct empirical finding. The picture that emerges is of a field with real momentum in its newest part and thin evidence overall, where the main task ahead is to move from architecture proposals to systems that are tested on real farms over more than one season.