Can an AI Agronomist Actually Beat a Farmer's Calendar?

Most agricultural advisories in India — the Package of Practices every state agriculture department publishes — get written before the growing season starts and then sit fixed for months. They tell a farmer when to sow, how much fertilizer to use, and when to irrigate, based on what usually works in that region. The problem is "usually." A fixed sowing date issued in March has no way to know that June will bring a late-season drought, and by the time a farmer feels the heat, the crop is already in its most vulnerable stage.

There's a newer layer of advice — in-season updates from India's Agro-Meteorological Field Units — but these are written by hand, one region at a time, and can't scale to every farmer's specific field. Meanwhile, a wave of AI chatbots has shown up promising personalized advice on demand. The catch: a language model doesn't actually know anything about crop physiology. It knows what sounds like good advice, which is a different and occasionally dangerous thing.

The IISc team's answer, called Agri-SAGE, is built around a specific discipline: before any recommendation reaches a farmer, it first has to survive a round with APSIM, a physics-based crop-growth simulator that's been used in agricultural research for over a decade. The AI proposes a plan — a sowing date, a fertilizer schedule, an irrigation amount — APSIM simulates what would actually happen to a real crop under that plan, and the AI has to look at the result (yield, water stress, nitrogen stress) and either defend its choice or revise it.

The researchers tested three different ways of managing that back-and-forth. One (Plan-and-Solve) diagnoses what went wrong after each attempt and revises. A second (Tree of Thoughts) sketches out three different strategies in advance, argues the pros and cons of each against the weather forecast, and commits to the best one before ever touching the simulator. A third (Reflexion) keeps a running memory of what worked in past seasons and leans on that experience instead of re-deriving everything from scratch each year.

To test this for real, the team didn't just run one clean season — they replayed ten years of actual historical weather (2015–2024) for Mandya, Karnataka, a region picked specifically because it swings between severe drought and heavy monsoon flooding. Every AI strategy had to plan a full maize-growing season against each of those ten years, compared against a "gold standard" baseline: the actual fixed advisory schedule real farmers in that region are given.

The baseline held up fine in ordinary years but fell apart exactly when it mattered — a 7,022 kg/ha yield in the 2019 drought, well below its own ten-year average. That's not a flaw specific to this one advisory; it's the structural weakness of any fixed calendar. It can't see a drought coming.

All three AI strategies beat the static baseline across the full decade, which is itself the headline finding. But they didn't all win the same way. Tree of Thoughts — the one that sketches multiple strategies and argues them out before committing — pulled ahead of the other two specifically during the worst years. In the 2019 drought, it looked at the forecast, decided mid-June sowing was going to fail, and simply moved sowing a full month earlier to May 15th, finishing the crop's growth before the worst heat arrived. It needed exactly one attempt to find that answer.

The reactive strategy (Plan-and-Solve) got the same result eventually, but only after failing first, receiving the bad news from the simulator, and course-correcting mid-season — the agricultural equivalent of learning from a mistake you didn't have to make. The memory-based strategy (Reflexion) sat in between: not as sharp in its first drought, but it got faster and better every subsequent year as its memory of past seasons filled in, needing fewer and fewer simulator attempts over time.