Look Ahead, Look Back, or Fix It Later: Three Ways to Build an AI Agronomist

TL;DR for operators

Agri-SAGE replaces the usual “retrieve some documents and produce a confident paragraph” workflow with a closed loop: retrieve locally relevant agronomic knowledge, generate a complete management plan, execute that plan inside the APSIM crop simulator, inspect yield and crop-stress signals, and revise the advice.

Within a ten-year retrospective maize simulation, all three tested reasoning strategies beat a static regional Package of Practices. Tree of Thoughts achieved the highest reported average simulated yield: 9,262 kg/ha, compared with 8,110 kg/ha for the static baseline. Plan-and-Solve reached 9,045 kg/ha, while Reflexion reached 9,002 kg/ha.

The strategies are not interchangeable. Tree of Thoughts spends more inference on comparing alternatives before acting. Plan-and-Solve commits to one plan, observes why it failed, and repairs it. Reflexion stores lessons from earlier simulated seasons so that recurring conditions can be handled with fewer new simulator interactions.

For businesses, the important result is architectural. When a credible domain simulator exists, an LLM can be used as a proposal generator rather than an unlicensed oracle. The simulator becomes a pre-deployment critic. That is more useful than ordinary retrieval augmentation, although considerably less magical than the phrase “autonomous AI agronomist” may suggest.

The boundary is equally important: every reported yield is produced by APSIM, not harvested from a field. The study does not yet establish profitability, field safety, environmental benefit, forecast robustness, or reliable performance outside Mandya maize.

Static instructions work beautifully until conditions change

A farming handbook can be scientifically sound and still give the wrong answer this year.

Regional Packages of Practices specify sensible sowing windows, fertilizer schedules, irrigation routines, and crop-management techniques. Their weakness is not ignorance. It is immobility. A calendar written before the season cannot move a crop’s reproductive stage away from an unexpected drought. It cannot delay nitrogen application because heavy rain is about to wash it away. It cannot reconsider a standard irrigation schedule after discovering that the soil is retaining less water than expected.

Dynamic agricultural advisories are intended to close this gap, but producing them manually requires experts who can interpret local weather, soil conditions, crop stages, and operational constraints at sufficient frequency. The scaling problem is obvious. So, naturally, the technology industry has proposed a chatbot.

Agri-SAGE takes the more defensible route. The system described by Balasubramaniam and colleagues does use an LLM, but it does not treat fluent language as evidence that an agronomic plan will work.¹ It connects the model to APSIM, a process-based agricultural simulator that models crop development, soil water, nutrient dynamics, biomass, and related physiological processes.

The practical workflow is:

Regional evidence + weather + soil → proposed season plan → APSIM simulation → yield and stress signals → revised plan

That final arrow changes the nature of the system. The model is no longer being asked merely to produce an answer that resembles expert advice. It is being asked to produce an executable plan whose consequences can be inspected inside a formal environment.

This does not make the answer true. It does make the answer testable.

Four contenders enter the simulator

The experiment compares a static baseline with three agent reasoning strategies. They share the same broad retrieval and verification architecture, but differ in when they spend their intelligence: before an action, after a failure, or across repeated seasons.

The static baseline is the region’s “gold standard” Package of Practices, translated into fixed APSIM operations. It specifies predetermined sowing dates, fertilizer splits, and other actions. The same management logic is then simulated under each year’s weather.

Plan-and-Solve is reactive. The agent proposes one plan and sends it to APSIM. It receives the simulated yield, daily water-stress factor, and nitrogen-stress factor. It then writes a critique of the previous plan, identifies a likely physiological failure, and generates a corrected version. The method is essentially structured trial and error, although “closed-loop biophysical optimization” looks better in a grant application.

Tree of Thoughts moves more reasoning in front of execution. For each decision stage, the agent generates three mutually exclusive agronomic strategies, critiques the advantages and disadvantages of each against the weather context, selects one path, and sends only that plan to APSIM. Instead of repairing a bad first idea, it attempts to avoid committing to one.

Reflexion adds cross-season memory. After a simulated season, the agent distils the outcome into textual heuristics, such as a warning that a particular sowing window performed poorly under drought conditions. Those lessons are inserted into the context for future years. The model weights do not change; the system improves by accumulating a written operating history.

These are three different approaches to uncertainty:

• Plan-and-Solve asks, “What went wrong?”
• Tree of Thoughts asks, “What else could we do?”
• Reflexion asks, “Have we seen something like this before?”

The paper’s most useful contribution is the controlled comparison among them.

The experiment tests simulated adaptation, not general agricultural intelligence

The authors evaluate maize in Mandya, Karnataka, using daily ERA5 weather data from 2015 through 2024 and a red sandy Alfisol soil profile. The period includes drought conditions and heavy monsoon years, creating substantial weather variation within one location.

The retrieval corpus contains roughly 1,000 open-access papers on maize cultivation, nitrogen efficiency, and drought mitigation, supplemented with official Karnataka extension material. A 4-bit quantized DeepSeek-R1 model, run at a temperature of 0.3, generates structured tool calls for APSIM. Those calls can alter sowing date, fertilizer type, fertilizer quantity and timing, irrigation, mulching, tillage, and related operations.

Each architecture, feedback-depth, and year combination is run ten times. The paper reports mean outcomes to reduce the effect of stochastic LLM generation.

There is an important experimental detail: the agents receive the static baseline yield as an optimization threshold. They are not independently deciding what success means. They are explicitly searching for a season plan that exceeds the Package-of-Practices outcome in the same simulator.

That is a legitimate optimization setup. It simply means the experiment evaluates whether an LLM agent can navigate APSIM’s management-action space more effectively than a fixed schedule. It does not test whether the agent can formulate the business objective, detect an unknown agronomic problem, or operate without a simulator-defined score.

Lookahead wins the decade-wide comparison

Across the reported ten-year means, each adaptive strategy beats the static Package-of-Practices schedule.

The order matters less than the reason for it.

Plan-and-Solve and Reflexion converge at similar average yields despite taking different routes. Plan-and-Solve uses current-season simulator feedback to recover from its own mistakes. Reflexion carries lessons across seasons and starts later years with a better prior. Both are primarily single-path strategies: they formulate one plan and then improve it through correction or memory.

Tree of Thoughts performs better because it searches across qualitatively different plans before committing. Its advantage is not simply that it “reasons more.” It is that the additional reasoning is organized around decision diversity. The model must consider multiple intervention pathways rather than polishing its first suggestion until the prose becomes reassuring.

That distinction becomes especially valuable when an early management choice changes the entire season. A suboptimal fertilizer split can sometimes be repaired. A sowing date that places flowering inside a drought window is more difficult to rescue. Lookahead is valuable when the system must avoid entering a bad trajectory, not merely adjust after one has begun.

The paper reports that Tree of Thoughts exceeds the baseline by 1,384 kg/ha during the 2019 drought and by 1,375 kg/ha during the heavy-rainfall conditions of 2024. In both cases, the relative improvement is roughly 19% over the reported static baseline for that year.

That is the core empirical argument: branching before execution appears most valuable under climatic stress, where choosing the wrong management trajectory creates larger downstream penalties.

There are, however, no reported confidence intervals, variance distributions, or formal statistical tests. The paper describes the gains as significant, but the available evidence demonstrates consistent differences in reported means rather than statistical significance in the conventional sense. Useful result, slightly enthusiastic adjective.

The drought example shows what each strategy actually does

The decade-wide averages establish the ranking. The 2019 drought analysis explains the mechanisms.

The static Package of Practices places sowing in mid-June. Under the simulated 2019 weather, that schedule leaves the crop’s sensitive reproductive stage exposed to a late-season drought. The resulting baseline yield is 7,022 kg/ha.

Plan-and-Solve initially follows a conventional schedule and encounters high water stress. APSIM exposes the problem. The agent then adds surface mulch to reduce evaporation and changes fertilizer selection to suit dry-soil conditions. In the qualitative case table, the revised plan reaches 8,168 kg/ha.

This is reactive intelligence. It does not avoid the original mistake, but it can diagnose and partially repair the consequences.

Tree of Thoughts compares alternative sowing strategies before execution. It selects an earlier date—May 15—so that the crop completes critical development before the harshest drought period. The method does not try to manage the drought better. It changes the timing so that the crop encounters less of it.

This is the clearest example of why lookahead outperforms correction. Once the model is required to compare genuinely distinct plans, it can discover that the best intervention is to avoid the predicted failure state rather than optimise within it.

Reflexion uses lessons retained from previous simulated seasons. Its plan applies 5,000 kg/ha of organic manure and deep tillage to improve soil water retention. The qualitative table reports 7,988 kg/ha, achieved on the first attempt.

This result is more nuanced than the paper’s broad “memory anticipates drought” framing. The authors also state that the system had not previously experienced a comparable drought and therefore required simulator feedback during the anomaly. Memory improved its starting position, but unusual conditions still caused simulator usage to rise from one iteration back to four.

That is exactly what should happen in a sensible memory system. Reuse is efficient when the environment resembles previous cases. Novel conditions should reduce confidence and trigger renewed testing. A memory mechanism that never notices novelty is not learning; it is accumulating habits.

The exact drought-year yields should nevertheless be treated cautiously. The qualitative table reports 8,825 kg/ha for Tree of Thoughts, while the main quantitative discussion says it exceeded the 7,022 kg/ha baseline by 1,384 kg/ha, implying approximately 8,406 kg/ha. Plan-and-Solve’s qualitative value also does not cleanly match its plotted decade trajectory. The paper may be comparing a representative run with a ten-run mean, but it does not explain the difference.

The mechanism is informative. The precise case-study magnitude is less settled.

The feedback-depth experiment is an ablation, not a second result

The study also varies the number of simulator feedback iterations. This is best read as an ablation of the closed loop: how much does repeated simulation add after the model’s initial retrieval-grounded proposal?

Plan-and-Solve shows the largest early improvement. Its first simulator response reveals obvious water and nutrient failures, allowing the model to capture low-hanging corrections. Gains then flatten after roughly three to four iterations.

Tree of Thoughts begins much closer to its eventual peak because it has already considered three alternatives before its first simulator action. Its first-step average is above 9,000 kg/ha in the convergence figure, and later loops add comparatively little.

Reflexion starts lower than the other two in the aggregate convergence plot, then improves rapidly after feedback. Across the year-by-year narrative, however, its first-iteration performance improves as the memory bank fills. The apparent tension is understandable: the aggregate figure mixes early memory-building seasons with later seasons in which stored experience is already useful.

The operational point is straightforward. More loops do not produce linear gains.

The paper describes feedback depths of zero, two, three, and four loops, while its convergence figure displays five numbered simulator steps. This may be an indexing difference between the initial execution and subsequent feedback loops, but it is not clarified. The broad plateau is visible; the exact loop-accounting convention is not.

Tree of Thoughts also consumes $k$ times more tokens at each decision stage because it generates $k=3$ branches. That indicates a real inference trade-off, but it should not be translated casually into “three times the total system cost.” APSIM execution, retrieval, translation, memory handling, and the number of subsequent loops also contribute to end-to-end cost. The study does not report wall-clock time, total token consumption, or monetary cost for each architecture.

The efficiency conclusion is therefore directional:

• Tree of Thoughts spends more language-model computation before acting.
• Plan-and-Solve spends more simulator interaction correcting a single plan.
• Reflexion can amortise earlier simulator work across later, similar seasons.

The final architecture choice depends on which of those resources is expensive.

The simulator turns the LLM into a proposal engine

Most domain-specific LLM systems use retrieval to improve factual grounding. Retrieval can tell the model which fertilizer types, sowing windows, and cultivation practices are documented for a region. It cannot establish that a particular combination will perform well under a particular sequence of weather and soil conditions.

Agri-SAGE adds a different form of grounding: consequence grounding.

The retrieval layer constrains the space of plausible actions. The simulator evaluates what those actions do. The feedback loop helps the model connect a management choice to a physiological outcome.

This architecture has a useful governance implication. The LLM does not have to be trusted as the final evaluator of its own recommendation. Its job is to propose, translate, and revise. A separate model of the domain supplies the outcome signals.

That separation is valuable wherever fluent recommendations are cheap but errors are expensive.

It is also fragile wherever the simulator is wrong.

The business decision is where to spend deliberation

Cognaptus’ inference from the study is broader than agriculture, but narrower than “agents can optimise complex systems.”

The reusable pattern is:

1. Restrict proposed actions using trusted domain knowledge.
2. Translate the proposal into a machine-executable plan.
3. evaluate the plan in a calibrated simulator.
4. Return a small number of actionable failure signals.
5. Decide whether to branch, repair, or reuse memory.

A business should choose the reasoning strategy according to the structure of its decisions.

Use lookahead when early choices are difficult to reverse

Tree of Thoughts is attractive when the first commitment determines much of the later outcome: production scheduling, supply-chain routing, capacity allocation, maintenance shutdown planning, or other problems where an early decision can trap the system in an expensive trajectory.

Its value comes from generating strategic diversity before execution. This is most defensible when the cost of a wrong first move exceeds the cost of examining several alternatives.

Use iterative repair when failures are observable and reversible

Plan-and-Solve fits settings where the system can run a proposal, receive informative diagnostics, and alter the plan without excessive switching cost.

It requires less parallel generation than Tree of Thoughts and does not need an established memory base. It is the sensible default for a new simulator-grounded workflow: one plan, one structured critique, one revision, then stop when improvements flatten.

Use memory when the organisation repeats similar decisions

Reflexion is potentially attractive for recurring planning cycles. A logistics network sees similar seasonal congestion. An industrial facility repeatedly handles related operating conditions. A farm encounters weather patterns that are not identical but share agronomically meaningful features.

In these settings, previous simulator outcomes can become a compact playbook that reduces repeated search.

The catch is memory governance. The stored lesson must include enough context to indicate when it applies. “Early sowing worked” is not a reusable rule. “Early sowing reduced reproductive-stage drought exposure under a particular rainfall and temperature pattern” is closer. Memory without conditions merely automates overgeneralisation.

Use none of them when the simulator is decorative

A simulator-grounded agent is only as grounded as the simulator.

Businesses should not insert a simplistic model into the loop merely to create the appearance of verification. The simulator must represent the variables affected by the proposed actions, expose useful failure diagnostics, and be calibrated well enough for the intended decision.

Otherwise, the workflow becomes an elaborate method for making the LLM agree with another model’s mistakes.

APSIM is a critic, not a certificate

The likely misconception is that simulation validation makes agricultural advice safe for field deployment.

It does not.

APSIM tests whether a management plan performs well under its encoded representation of crop physiology, weather, soil, and management operations. This is stronger evidence than an LLM’s linguistic confidence. It is not equivalent to observing the plan across real farms.

The comparison is internally controlled because the static baseline and all three agent strategies are evaluated inside the same simulator. This reduces noise when comparing reasoning methods. It also means every approach shares APSIM’s assumptions and blind spots.

The experiment does not incorporate several variables that would matter immediately to an operator:

• Input prices and farmer cash constraints
• Labour and equipment availability
• Local access to specified fertilizers or organic manure
• Pest and disease dynamics
• Environmental impacts such as nitrogen leaching
• Carbon emissions
• Profit volatility
• Weather-forecast errors
• Compliance with local agronomic or chemical-use requirements
• Variation among actual fields within the region

Yield is treated as the primary economic objective, but yield is not profit. Applying more fertilizer, manure, irrigation, tillage, or mulch may increase simulated output while reducing the farmer’s margin or creating environmental costs.

A business deployment would therefore need a multi-objective score rather than one final yield number. At minimum:

The coefficients are not technical details. They encode whose interests the advisory system optimises.

The evidence ends at the edge of the simulated field

The paper provides a promising architecture demonstration. It does not yet provide a deployment case.

First, the study covers one crop, one regional soil profile, and ten historical weather years. The architecture may be crop-agnostic in software terms, but empirical transfer to another crop requires an adequate APSIM model, different management actions, appropriate regional knowledge, and new validation.

Second, the agents are evaluated using retrospective ERA5 weather. A production advisory system would make decisions using forecasts, which contain timing and magnitude errors. Tree of Thoughts may be especially sensitive to this distinction because its advantage comes partly from selecting a trajectory based on anticipated conditions. Looking ahead helps only when the view is reasonably accurate.

Third, the reported outcomes are simulator predictions. No field trials compare agent-generated plans with standard practice. Field heterogeneity, equipment constraints, farmer behaviour, and unmodelled biological processes may change both the ranking and the magnitude of the results.

Fourth, the statistical reporting is limited. Ten independent LLM runs are used, but the paper does not show dispersion, confidence intervals, or significance tests. Internal inconsistencies between the drought table and the main quantitative discussion further weaken confidence in exact case-level values.

Fifth, the computational comparison remains incomplete. Tree of Thoughts clearly generates more candidate text, and Reflexion appears to reduce simulator usage in familiar later seasons. But there is no complete accounting of token consumption, simulator runtime, latency, memory overhead, or financial cost.

These are not reasons to dismiss the work. They define the next experiment.

A convincing follow-up would compare the strategies prospectively using imperfect weather forecasts, include profit and environmental objectives, report complete cost and uncertainty statistics, and run field or at least multi-location validation. It should also test whether memory learned in one sequence of seasons transfers safely when the climate regime changes.

The useful AI agronomist is the one that can be contradicted

Agri-SAGE’s strongest idea is not that an LLM can generate better farming advice. LLMs can generate advice about practically anything, which has never been the reassuring part.

The stronger idea is that an LLM’s proposal can be converted into actions, executed in a structured representation of the domain, criticised using physical outcome signals, and rejected or revised before reaching a person.

The comparison among the three reasoning strategies then becomes a deployment choice rather than an abstract prompting contest.

Tree of Thoughts buys better simulated outcomes by spending more computation on alternatives. Plan-and-Solve offers a straightforward correction loop. Reflexion attempts to turn repeated simulator work into institutional memory. None is universally best. Each moves the cost of deliberation to a different place.

For agricultural advisory, that is progress—but it is pre-field progress. The system has learned to survive contact with a crop model. Surviving contact with crops, farmers, budgets, equipment, and weather forecasts remains the less convenient phase of the project.

Cognaptus: Automate the Present, Incubate the Future.