🤖 AI

Agri-SAGE: Simulation-Grounded Multi-Agent LLM for Context-Aware Agricultural Advisory Generation

Agri-SAGE is a closed-loop framework that integrates retrieval-grounded multi-agent LLM reasoning with APSIM biophysical simulations to generate and validate dynamic, context-aware agricultural advisories, demonstrating that reasoning strategies like Tree of Thoughts and Reflexion significantly outperform static guidelines in yield optimization while balancing performance with computational efficiency.

Original authors: Vedant Balasubramaniam, Geetha Charan, Manojkumar Patil, Rohit P Suresh, V Priyanka, Kodur Sai Vinay Sathvik, Y. Narahari

Published 2026-07-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Vedant Balasubramaniam, Geetha Charan, Manojkumar Patil, Rohit P Suresh, V Priyanka, Kodur Sai Vinay Sathvik, Y. Narahari

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guide a farmer through a growing season. Traditionally, you'd hand them a static instruction manual (called a "Package of Practices"). This manual is like a printed cookbook written years ago: it says, "Plant corn on June 15th and add fertilizer this way." It's based on good science, but it's blind. If a drought hits in July or a flood comes in August, the manual doesn't know. The farmer follows the instructions anyway, and the crops suffer.

On the other hand, you could ask a super-smart AI chatbot (a Large Language Model or LLM) for advice. This AI has read millions of books and knows a lot about farming. But here's the catch: the AI is like a brilliant theorist who has never actually planted a seed. It might give you a recommendation that sounds perfect in English but is physically impossible in the real world (like telling a plant to drink water when the soil is already a swamp).

Agri-SAGE is a new system designed to fix both problems. Think of it as a team of three specialized AI farmers working together with a virtual reality simulator to create a perfect, custom plan for the farmer.

The Team: How Agri-SAGE Works

The system is a "closed-loop" team, meaning they talk to each other, make a plan, test it, and fix it before telling the farmer anything.

The Librarian (Retrieval Agent): Before making any suggestions, this agent looks up the farmer's specific local rules, soil type, and the latest weather forecast. It makes sure the advice is grounded in reality, not just random guesses.
The Strategist (Generation Agent): This is the main brain. It takes the info from the Librarian and the current weather to create a farming plan (when to plant, how much water to give, etc.).
The Simulator (APSIM): This is the magic trick. Before the Strategist tells the farmer what to do, it runs the plan through a high-tech video game called APSIM. This simulator acts like a "physics engine" for crops. It calculates: If we do this, will the corn actually grow? Will it die from lack of water?
The Critic (Verification Agent): This agent looks at the simulator's results. If the virtual corn is dying of thirst, the Critic tells the Strategist, "Hey, that plan failed. Try again." The Strategist then rewrites the plan and runs it through the simulator again until it works.

The Three Thinking Styles

The researchers tested three different ways for the "Strategist" to think, similar to how a human might solve a puzzle:

Plan-and-Solve (The "Try and Fix" Approach):
- Analogy: Imagine you are cooking a new dish. You taste it, realize it's too salty, add some sugar, taste it again, and adjust.
- How it works: The AI makes a plan, runs it in the simulator, sees the result, critiques its own mistakes, and fixes the plan. It repeats this loop until the crop yield is high.
Tree of Thoughts (The "What-If" Explorer):
- Analogy: Imagine you are at a fork in the road. Instead of just picking one path, you imagine walking down three different paths simultaneously. You visualize where each path leads, pick the best one, and then walk it.
- How it works: Before making a move, the AI generates three different farming strategies. It simulates all three, compares them, and picks the absolute best one. This prevents the AI from getting stuck on a bad idea early on.
Reflexion (The "Memory Keeper"):
- Analogy: Imagine a farmer who keeps a diary. Last year, they learned that planting in June during a drought was a disaster. This year, they read their diary and remember, "Don't plant in June!"
- How it works: The AI keeps a text-based memory of past seasons. If it sees a drought coming, it remembers, "Oh, I've seen this before; I need to change my strategy." It learns from its past "virtual" failures without needing to run the simulator as many times.

What Happened in the Test?

The researchers tested this system in Mandya, India, using 10 years of real weather data (including severe droughts and heavy rains) on Maize (corn). They compared their AI team against the traditional "Static Manual."

The Result: The traditional manual gave an average yield of about 8,110 kg/ha.
The AI Team: All three AI methods beat the manual.
- Tree of Thoughts was the champion, achieving the highest yields (about 9,262 kg/ha). It was the best at avoiding disasters by looking ahead and choosing the best path before starting.
- Plan-and-Solve and Reflexion also did very well, beating the manual by a significant margin.
- Reflexion was the most efficient. Because it learned from its "diary" of past seasons, it didn't need to run as many simulations to get a good result, saving computer power.

The "2019 Drought" Story

To see how they handled a crisis, the researchers looked at the year 2019, which had a terrible late-season drought.

The Manual: Stuck to the calendar. It told farmers to plant in mid-June. The corn grew, then the drought hit right when the corn needed water most. Result: Low yield.
Plan-and-Solve: Tried the standard plan, saw the corn dying in the simulator, and said, "Okay, let's try adding mulch (a layer of straw) to keep moisture in and change the fertilizer." Result: Better yield.
Tree of Thoughts: Before planting, it simulated three options. It realized, "If we plant in June, the drought will kill us. But if we plant in May, we'll finish growing before the heat hits." It changed the planting date entirely. Result: Best yield.
Reflexion: Remembered that dry soil holds water poorly. It added a huge amount of organic manure to the soil to act like a sponge, holding more water for the dry spell. Result: High yield, achieved quickly.

The Bottom Line

Agri-SAGE proves that you can't just rely on a static rulebook or a chatbot that only knows words. You need a system that thinks, simulates the physical reality of the farm, and learns from its mistakes.

By combining the knowledge of an AI with the "physics engine" of a crop simulator, the system creates farming advice that is not only smart but also physically possible and safe to follow, even when the weather is crazy.

Technical Summary: Agri-SAGE

Problem Statement

Agricultural advisory systems face a critical dichotomy. Traditional Packages of Practices (PoPs) offer static, evidence-based guidelines derived from historical field trials but fail to adapt to in-season variability, such as weather shocks, pest outbreaks, or localized soil heterogeneity. Conversely, emerging Large Language Model (LLM)-based advisory systems improve accessibility and knowledge synthesis but suffer from a different limitation: they may generate recommendations that are agronomically plausible in text but physiologically unconvincing or unsafe to act upon. Current LLM systems often lack a mechanism to validate their reasoning against biophysical reality, leading to "hallucinated" advice that does not account for crop physiology or environmental constraints.

Methodology: Agri-SAGE Framework

Agri-SAGE (Simulation-Grounded Multi-Agent LLM for Context-Aware Agricultural Advisory Generation) is a closed-loop framework designed to resolve these limitations by integrating retrieval-grounded multi-agent LLM reasoning with the APSIM (Agricultural Production Systems sIMulator) biophysical crop model.

The architecture consists of three primary modules:

Retrieval Agent: Grounds the system in a curated knowledge base of regional manuscripts and PoPs. It uses semantic-aware hierarchical chunking and hybrid dense-sparse retrieval to process real-time environmental inputs (weather, soil, phenology) and retrieve context-specific agronomic literature.
Generation Agent: The central decision-making component that synthesizes retrieved agronomic context with real-time environmental states. It formulates actionable advisories (e.g., sowing dates, fertilizer splits, irrigation schedules) and is evaluated using three distinct reasoning methodologies:
- Plan-and-Solve (PS): An iterative closed-loop approach where the agent generates a textual critique of biophysical failures from the previous step, revises the plan, and executes the new action.
- Tree of Thoughts (ToT): An anticipatory lookahead strategy where the agent generates $k=3$ mutually exclusive intervention pathways, explicitly critiques their biophysical pros and cons against current forecasts, and selects the highest-scoring path before execution.
- Reflexion: An approach integrating a persistent episodic memory module. It stores distilled heuristic rules from past simulation years (e.g., "mid-June sowing fails in drought years"), allowing the agent to condition its policy on historical context for zero-shot adaptation.
Verification and Feedback Agent: Translates LLM outputs into structured tool calls for APSIM. It simulates the full crop life cycle and extracts three key feedback metrics: Grain Yield (primary optimization objective), Water Stress Factor, and Nitrogen Stress Factor. These metrics allow the LLM to iteratively critique and refine its advisories based on physiological drivers of crop success or failure.

Experimental Setup

The framework was evaluated over a 10-year retrospective analysis (2015–2024) in Mandya, Karnataka, India, a semi-arid region characterized by high weather variability.

Crop: Maize (Zea mays).
Data: Daily weather data from the ERA5 dataset and Mandya red sandy soil (Alfisol) profiles.
Baseline: A static "Gold Standard" PoP schedule translated into fixed APSIM tool calls.
Model: A 4-bit quantized DeepSeek-R1 model with a generation temperature of 0.3 to ensure deterministic logic while retaining exploration capability.
Protocol: Each configuration was executed in 10 independent runs to stabilize statistical variance. An ablation study was performed on feedback depths ( $N \in \{0, 2, 3, 4\}$ ).

Key Results

All three reasoning approaches significantly outperformed the static PoP baseline, which achieved a 10-year average yield of 8,110 kg/ha.

Tree of Thoughts (ToT): Achieved the highest performance with a 10-year average yield of 9,262 kg/ha (+1,152 kg/ha over baseline). Its superiority stems from its anticipatory design, which avoids failure modes common to reactive strategies by evaluating multiple pathways before execution. During extreme events (e.g., the 2019 drought), ToT outperformed the baseline by 1,384 kg/ha by shifting sowing dates to avoid stress entirely.
Plan-and-Solve (PS) & Reflexion: Both achieved comparable yield ceilings (~9,000 kg/ha), outperforming the baseline by approximately 900 kg/ha.
- PS improved through iterative self-correction based on simulator feedback.
- Reflexion leveraged cross-seasonal episodic memory to achieve competitive yields with substantially lower computational cost. It demonstrated faster convergence, often reaching optimal yields in the first iteration by applying learned heuristics, whereas ToT consumes more tokens due to parallel path exploration.

Convergence Dynamics: The study observed that forcing agents beyond 3–4 feedback loops yields diminishing returns. While ToT achieves peak yields, it is computationally heavier; Reflexion and PS offer a more efficient trade-off by incrementally refining a single path.

Significance and Contributions

The paper claims three primary contributions:

Novel Framework: Agri-SAGE introduces a closed-loop autonomous AI Agronomist that successfully couples LLMs with biophysical crop simulation, grounding generative semantic reasoning in crop physiology rather than solely in text.
Rigorous Evaluation: The study provides a comparative analysis of advanced reasoning methodologies (PS, ToT, Reflexion) in an agricultural context, demonstrating that all significantly outperform static baselines.
Discovery of Optimized Practices: The framework discovers significantly better agronomic practices than static PoPs. For instance, in the Mandya region, the system optimized fertilizer splits, irrigation timing, and operations like surface mulching to improve outcomes under local conditions.

The authors conclude that coupling LLM reasoning with crop simulation transforms the state of the art in agronomic advisories, moving from static, one-size-fits-all recommendations to dynamic, context-aware, and biophysically validated management strategies. They note that future work should extend optimization objectives to include economic costs and sustainability metrics, incorporate real-time weather forecasts, and pursue field validation to address real-world soil and microclimate variability.