Source-linked AI summary
Deploying and Evaluating a Smart-Agriculture Agentic Engine for Full-Season Soybean Farm Operations
Ao Qu, Panagiotis Michelakis, Linyuan Han, Yiannis Hadjiyianni, Kun Ouyang, Konstantinos Siskos, Feng Li, Ran Meng, Jingchi Jiang, Dimitrios Stamoulis, Jie Liu
TL;DR
Agricultural agents must operate under delayed feedback, partial spatial observations, and consequences that unfold across biological and operational processes. FAIRY addresses this gap with an event-driven full-stack engine for a soybean research farm and evaluates nine controllers across 100 realistic scenarios. Expert context and retrievable agronomic skills improve long-horizon reliability, while orchestration cost and digital-twin evaluation define important deployment boundaries.
Problem
Agricultural agent systems need evaluation that captures spatial coverage, temporal alignment, physical crop processes, delayed effects, and operational meaning beyond successful tool calls.
Method
FAIRY integrates farm infrastructure, sensing, crop-process models, agronomic knowledge, event-based state transitions, controller backends, trace logging, and 100 multi-complexity soybean scenarios.
Results
Expert operational context reduces full-season yield shortfall from ∼22% with zero context to ∼5% for Qwen on held-out L3 scenarios.
Takeaways & Limitations
Expert context and retrievable agronomic skills are dominant levers for long-horizon reliability, while multi-agent orchestration and deployment cost remain practical constraints.
Takeaways & Limitations
The study is a deployment-readiness evaluation: agent actions are replayed through a calibrated digital twin rather than a completed agent-managed harvest trial.
Abstract
from arXiv · showhide
This paper presents FAIRY, a full-stack smart-agriculture agent system developed for and deployed to an operating soybean research farm at Harbin Institute of Technology's smart-agriculture site. We develop FAIRY to execute and evaluate agentic agronomic operations on full-season spatiotemporal workflows that span ridge preparation, planting, irrigation, fertilization, pest and disease treatment, harvest, grain handling, drying, and storage. FAIRY integrates APIs and infrastructure across production-grade machinery, fixed soil and canopy sensors, multispectral and thermal drones, satellite vegetation products, a weather station, calibrated crop-process models, agronomic records, and multi-season yield histories. The system is built around the novel "everything is an event" execution paradigm, which represents spatiotemporal world evolution, remote sensing and UAV observations, sensor readings, crop-growth transitions, machinery actions, and management interventions as state-changing events in a shared farm process engine. On top of this event-driven world model, FAIRY implements a complete agentic stack: a knowledge library of atomic agronomic skills; multi-agent controller and orchestration backends; frontier- and edge-model execution; full-path trace logging; and deployment profiling on local nodes. We use FAIRY to evaluate nine state-of-the-art agent controllers across one hundred full-season soybean scenarios that preserve the operational coupling between spatial observations in a 64-ridge field, temporal decision sequences, agronomic constraints, delayed effects, and final yield. We develop an evaluation suite that combines agentic success, full-path spatiotemporal correctness, token cost, and edge-device runtime.
1 Introduction
FAIRY addresses the difficulty of deploying agents for agricultural workflows with delayed feedback, partial spatial observations, and compounding consequences. It provides a full-stack, event-driven research engine and evaluates nine controllers across realistic full-season soybean scenarios.
- Motivation: Agricultural operations require reasoning over delayed feedback, partial spatial observations, and consequences that propagate across the season.These conditions can make syntactically valid actions operationally wrong, such as harvesting before grain moisture is suitable.
- System and setting: FAIRY integrates farm machinery, sensing infrastructure, crop-process models, agronomic records, and multi-season yield histories for an operating 64-ridge soybean research farm.The farm workflow spans ridge preparation, planting, irrigation, fertilization, treatment, harvest, grain handling, drying, and storage.
- System and setting: FAIRY represents weather, observations, sensor readings, crop transitions, machinery actions, and interventions as state-changing events in a shared farm process engine.Its agentic stack includes atomic agronomic skills, nine controller backends, optional A2A orchestration, frontier- and edge-model execution, trace logging, and deployment profiling.
- Evaluation: One hundred scenarios cover atomic tasks, episode chains, and full-season workflows while preserving field geometry, crop-process assumptions, sensing constraints, and operational coupling.The evaluation suite measures task success, temporal correctness, coverage, full-path correctness, and yield preservation.
- Findings: Expert operational context reduces the full-season yield shortfall from ∼22% under zero context to ∼5% for Qwen on held-out L3 scenarios.The results also report that longer-horizon scenarios remain materially below the human oracle and that A2A coordination can degrade correctness and cost.
2 The FAIRY System: Architecture Overview
FAIRY connects a 64-ridge soybean research farm’s equipment and sensing assets to agent-callable interfaces, satellite and drone processing pipelines, and a broader event-driven system architecture.
- Architecture: The FAIRY architecture couples an event-driven engine, physics-grounded process stack, tool and sensing layer, knowledge library, and agentic backend.This overview frames the system as an integrated deployment and evaluation platform.
- Farm site and APIs: The modeled 268 m × 71 m field is organized into 64 ridges serving as atomic spatial units for observation and intervention.Agent function-calling tools expose installed farm APIs and operational interfaces.
- Farm site and APIs: FAIRY combines fixed environmental and crop sensors, UAV and calibration assets, ground robots, production machinery, spraying equipment, and ridge-level irrigation and fertilization facilities.These assets closely model the underlying farm equipment and on-field sensing infrastructure.
- Satellite products: The satellite pipeline filters Sentinel-2 imagery by area, date, and cloud coverage, then exports selected scenes as 10 m GeoTIFFs in EPSG:4326.The resulting raster is clipped and rendered as a Leaflet map backdrop using RGB bands.
- Satellite products: Satellite crop classification uses an in-house XGBoost multiclass model with ten Sentinel-2 bands and vegetation indices, trained with a stratified 70/30 split.The operational classes include rice, maize, and soybean, while regional fine-tuning is disabled in this study.
- Drone operations: Drone operations use DJI M300-RTK imagery, DJI Terra orthomosaics, optional OpenDroneMap processing, and UAV API experimentation for field inspection.The drone system produces 8192×5460 full-frame RGB imagery.
2.2 Growth process dynamics
FAIRY models soybean growth through coupled weather, soil, phenology, canopy, biomass, and biotic-pressure dynamics, linking management actions to delayed yield effects and historical calibration.
- Environmental drivers: Weather inputs provide precipitation, temperature, radiation, wind, and humidity to the farm event loop alongside remote-sensing and drone observations.The weather model follows a WGEN/Richardson-style daily formulation.
- Crop-process model: The growth stack combines bucket-style soil-water balance, GDD-based soybean phenology, canopy and biomass dynamics, and weather- and stage-dependent weed, insect, and disease pressure.Phenology supports seed-type-specific maturity targets and cultivar-specific assumptions.
- Management and yield: Management effects depend on crop stage, soil condition, recent weather, and canopy state, with planting, irrigation, and fertilization altering later crop trajectories.Planting affects stand fraction and emergence timing; irrigation changes subsequent soil-water availability; fertilization affects nutrient stress and canopy development.
- Calibration: Eighteen 2025 plot-level soybean scenarios calibrate FAIRY against observed phenology and yield records using plot-specific management, cultivar, density, fertilization, and weather conditions.The scenarios span planting dates, density treatments, and cultivar types.
2.3 Stateful event-driven engine
FAIRY uses a stateful event-driven engine in which farm tools execute timestamped operations with preconditions, preserving farm-state effects and timing dependencies in a calibrated digital twin.
- Event model: The execution engine represents farm environments through stateful applications, an event queue, notifications, scenarios, and logged execution traces.Operations such as planting, irrigation, pesticide application, harvest, and storage become timestamped farm events.
- Event model: Each farm event carries arguments and preconditions, while the environment maintains state, queued events, notifications, and operation history.Examples include operation duration, target ridges, weather readiness, and equipment readiness.
- Digital twin: Tool invocation is emulated inside the digital twin so operations retain farm-state effects and timing dependencies without real-clock execution.Controllers observe and act through the tool interface while the environment updates the shared farm state.
- Scope: The study evaluates deployment readiness rather than a completed agent-managed harvest trial because agent actions are replayed through the calibrated digital twin.The physical infrastructure, sensing stack, machinery interfaces, and historical operation records are real.
2.4 Agent families and LLMs
FAIRY evaluates nine controller architectures under consistent farm-state and tool conditions, with direct-tool and agent-to-agent modes. Its controllers can retrieve structured agronomic skills or operate under alternative context regimes, including human-written instructions.
- Controller architectures: FAIRY integrates nine controller architectures, each supporting direct tool access and agent-to-agent routing through specialist weather, sensing, machinery, and operations agents.The evaluated controllers are ReAct, Plan-and-Act, Reflexion, AutoGen, MMRL, ReWOO, LATS, CRITIC, and GoT.
- Agronomic context: FAIRY encodes tacit farm-management timing rules as retrievable agronomic skills with identifiers, descriptions, keywords, and reusable workflow templates.A retriever scores skills against the current task and injects the top-k entries at decision time.
- Context regimes: Controllers run under four context regimes: Zero Context, LLM-as-an-Expert, Skills Library, and Expert Instruct.The study also varies whether skills are organized as a flat pool or tier-grouped retrieval structure.
2.6 Realistic full-season scenarios
FAIRY constructs and evaluates increasingly complex soybean-operation scenarios, from short tool tasks to full-season workflows with delayed consequences. Agent traces are compared with human-oracle event trajectories using temporal correctness and yield outcomes.
- Scenario design: L1 atomic tasks use one or two tool calls, L2 episodes chain observation, diagnosis, and intervention, and L3 scenarios span planting through post-harvest storage.L3 scenarios include multiple interventions, delayed consequences, and accumulated effects on crop state and recovered yield.
- Scenario design: Example scenarios range from verifying aphids and applying pesticide treatment to diagnosing low NDVI and managing a full season with weekly monitoring.The examples preserve ridge-level spatial targeting and interventions across water, nutrient, pest, and disease issues.
- Evaluation sets: The evaluation uses 70 L3 test scenarios, a focused 20-scenario L3-mini ablation set, and 20 held-out validation scenarios for library-scheme generalization.Oracle workflows provide farm-state trajectories and target crop yields for comparison.
- Evaluation metrics: KTC measures normalized Kendall rank agreement between matched agent and oracle operations, rewarding causally valid order and penalizing reordering.KTC = (τ + 1)/2 ∈ [0, 1].
- Evaluation metrics: Table 2 compares Qwen3.6-35B-A3B and DeepSeek-V4-Flash across L1, L2, and L3 using Yield Loss, Succ., KTC, and token cost.Token cost is reported as average tokens per task and per agent API call.
- Evaluation metrics: Yield Loss measures the percentage shortfall from human-oracle biological yield, complementing trace correctness by capturing operational misjudgments.The evaluation also reports token cost and tool-call success alongside correctness and yield outcomes.
2.8 FAIRY system user interface
FAIRY adapts a generic agent-workflow interface into a geospatial inspection environment for agricultural operations. The interface combines workflow execution inspection with map-based satellite and UAV visualizations.
- Geospatial visualization: The interface adds zoomable OpenStreetMap views, satellite imagery overlays, adjustable-opacity crop-classification rasters, and UAV patrol and orthomosaic tabs.These visualizations support inspection of satellite and drone-analysis outputs within agricultural workflows.
- Workflow inspection: Dedicated state and tool panels let users inspect the execution chain, intermediate outputs, and final geospatial products in one workflow interface.The web-based viewer displays the full-path agent workflow alongside map-based visualizations.
3 Results
FAIRY results show that agronomic context and structured retrieval are central to reliable full-season behavior, while longer horizons and multi-agent coordination expose substantial weaknesses. The evaluation also links trace correctness to yield, profiles edge deployment, and identifies task-decomposition errors as future work.
- Context and skills: Expert instructions reduce Qwen Yield Loss from 22.3% to 4.6% and raise KTC from 87.9% to 92.1% on held-out L3 scenarios.The Skills Library achieves 4.9% Yield Loss and 93.7% KTC without hand-written per-scenario policies.
- Context and skills: Tier-grouped similarity retrieval reaches up to 97.2% KTC for Qwen and 96.8% for DeepSeek, with stronger task success than flat/manual retrieval.Retrieval improves correctness more than yield because it refines behavior after context grounding has already occurred.
- Horizon and coordination: Under expert context, L1 reaches ≥99% KTC with near-zero yield loss and L2 reaches ≥98% KTC with <1% yield loss, unlike materially weaker L3 performance.Full-season scenarios require coherence after effects propagate through soil, growth, stress, treatments, harvest timing, and storage.
- Horizon and coordination: A2A routing raises Yield Loss from 7.1% to 31.2% for DeepSeek and from 10.7% to 29.4% for Qwen, despite lower token cost per scenario.The main controller frequently loses track of specialist observations or actions, while tree search is comparatively robust.
- Evaluation and deployment: KTC is the best-calibrated yield predictor, whereas BFCL mainly detects catastrophic never-harvested runs and may be poorly calibrated among completing runs.The practical recommendation is to report KTC for trace correctness, Yield Loss for outcome, and exact-match success as a failure detector.
- Evaluation and deployment: Qwen3.6-35B-A3B-FP8 and Qwen3.6-27B-FP8 keep L3-mini yield loss near 2% under local vLLM execution, while smaller or assistant-tuned variants degrade sharply.The profiling results indicate edge deployment is feasible for evaluation, but full-season autonomy requires sufficient planning capacity.
4 FAIRY Operational Demonstration
FAIRY demonstrates an end-to-end field-monitoring episode in which an agent coordinates satellite and drone observations, analyzes crops, and updates farm-facing geospatial visualizations.
- The L2 episode begins with a field-monitoring request and coordinates satellite and drone observations.
- The agent triggers the drone API, retrieves satellite imagery, and obtains the drone result during the inspection workflow.
- The workflow generates an orthomosaic, plots the patrol path, analyzes the field and crop, and integrates satellite and drone findings.
- The deployment prototype connects a user request to event-driven agent execution and farm-facing geospatial visualization.
5 Discussion
The discussion emphasizes reusable agronomic skills, full-path outcome evaluation, state preservation across orchestration, and disciplined validation of human-oracle workflows.
- From expert instructions to reusable skills: Retrievable, composable agronomic skill libraries offer a scalable alternative to writing new expert instructions for every task or seasonal edge case.
- Full-path evaluation paired with farm objectives: KTC and Yield Loss should be evaluated together because locally plausible actions can still produce downstream yield loss.
- Agent orchestration should preserve operational state: Orchestration requires shared farm state, timing constraints, and task context to remain intact across specialist boundaries.
- LLM-as-an-Expert context: Agriculture-tuned LLM context recovers high-level crop-management choices but does not fully match human experts on procedural ordering.
- Quality of human-annotated solutions: Human-oracle workflows should be version-controlled and serialized because regeneration can silently change every trace metric.
- The next deployment stage compares agent-managed and human-operated plots using operational traces and harvested outcomes under real seasonal conditions.
6 Conclusion
FAIRY is a deployable smart-agriculture engine used to evaluate contemporary agent practices on full-season soybean operations. The conclusion identifies expertise and reusable skills as central to long-horizon reliability while highlighting orchestration and deployment constraints.
- FAIRY provides a deployable engine for full-season spatiotemporal evaluation of agent practices on an operating soybean research farm.
- Expert context and retrievable agronomic skills are identified as dominant levers for long-horizon reliability.
- Multi-agent orchestration and deployment cost remain practical constraints, while temporally grounded correctness tracks real yield among trace metrics.
- The next stage is to operate a dedicated agent-managed plot alongside the human-operated farm.