Future-proofing agricultural research: FAIR principles for agriculture AI agents (FAIR4AG²)

The rapid shift of large language models from conversational use to agentic reasoning is changing how scientific outputs must be structured for machine consumption. Agriculture stands to gain the most from this transition but currently has the least of the centralized, machine-ready infrastructure that biomedicine has built over decades. Agricultural knowledge remains dispersed across peer-reviewed journals, extension bulletins, technical reports, and multimedia field demonstrations, with associated code, data, and models often inaccessible to autonomous agents. We argue that the foundational FAIR principles and FAIR for Research Software (FAIR4RS) must be extended to a new standard of agent-actionability, and propose FAIR4AG² as that extension for agricultural research. Across three modalities of knowledge units (text, multimedia, and databases), we offer concrete, implementation-ready practices: structured publishing formats and machine-readable licensing for documents; signal isolation, time-aligned visuals, and domain-aware curation for multimedia; and standardized APIs, Agent Skills, and Model Context Protocol (MCP) servers for databases and model repositories. Realizing FAIR4AG² will require parallel investment in equitable participation, careful curation, human-in-the-loop verification, and governance norms that credit the data curators and infrastructure builders whose work agents now operate upon.