Researchers have built a system that makes AI farming advisors verify their own recommendations against real crop science before passing them on.
Called Agri-SAGE, the framework connects multi-agent large language models to APSIM, a well-established biophysical crop simulator, in a closed loop. The idea is to catch a specific failure mode: LLMs that produce advice sounding credible to an agronomist but that wouldn't survive contact with actual plant physiology. The team tested three reasoning strategies — Plan-and-Solve, Tree of Thoughts, and Reflexion — against a 10-year retrospective dataset. All three beat static "package-of-practice" baselines, the kind of fixed seasonal guidelines most advisory systems still rely on. Tree of Thoughts hit the highest peak yields; Reflexion matched it on agronomic outcomes while using substantially less compute by drawing on cross-seasonal memory rather than re-reasoning from scratch each time.
The gap this targets is real. Static guidelines can't adapt mid-season when weather turns or pest pressure shifts, and pure LLM systems will confidently recommend something that sounds right but ignores how a specific crop actually responds at a physiological level. Grounding recommendations in simulation output gives the system a falsification step that neither approach had before.
That said, APSIM models are only as good as their calibration data, and a simulator confidently wrong is still wrong — so the "closed loop" here is only as tight as the underlying biophysical assumptions.