Having explored some multi-agent AI workflows in subsurface analysis, I have some ground for understanding the basis for using multiple AI agents to optimize insights from data.
Machine Learning (ML), a
subset of AI, has been utilized in soil science to solve specific problems and
has proved useful in a predictive capacity. What ML-based soil science can’t do
is robust investigations of the underlying cause-and-effect mechanisms
resulting in the generation of hypotheses. The authors cite a study by Khanifar
that evaluated LLMs for soil science knowledge and “found that, on average,
they could correctly answer only up to 65% of questions from advanced soil
science examinations.” That is OK but certainly not good enough.
The authors note that
utilizing multi-agent AI in soil science should emulate successes in other
fields:
“An AI agent is a software system that uses AI to
interact with its information environment, gather information, and perform
tasks to achieve specific goals.”
“The vision of AI agents transforming scientific
discovery is gaining momentum in fields such as chemistry, materials science,
and biomedicine, where complex systems, vast datasets, and multidisciplinary
knowledge must be integrated.”
The goals of the study are listed below:
Ideally, an AI agent becomes a co-scientist, a partner. AI agents may have the ability to solve complex problems in complex domains like soil science. The figure below shows how five AI agents with specific tasks can be integrated in order to solve the problem of assessing the effects of climatic changes on soil organic matter retention.
Soil carbon storage is affected by soil heat, which increases microbial activity that breaks down organic matter faster, and by soil moisture.
Multi-agent AI has emerged as
a very effective means of advancing scientific research and can accelerate
discoveries.
“A multi-agent system can act as an orchestrator,
selecting, configuring, and executing these tools within a coordinated
workflow, moving beyond the current fragmented use of individual technologies.
The agents possess perceived reasoning and planning capabilities that extend
well beyond simple data processing.”
“These systems emulate collaborative scientific
workflows involving brainstorming, critique, consensus-building, and iterative
refinement.”
The table below summarizes
the key features of multi-agent AI models in scientific research. There are
four models: 1) general multi-agent AI systems, 2) AI co-scientists (which can
refine hypotheses), 3) astro agents (which can analyze large data sets, finding
clusters and utilizing pattern recognition), and 4) spatial agents used in
biomedical research.
These models share core
principles and capabilities:
Below, the authors propose a
multi-agentic system for soil analysis to be utilized for hypothesis generation
and experiment design. It integrates key components from leading multi-agent AI
models. Agents and sub-agents are given the goal of proceeding from knowledge
acquisition to hypothesis generation to evaluation and feedback, and to arrive
at a pool of hypotheses. A Meta-Review Agent then aggregates these evaluation
results and ranks the hypotheses, which are then selected and evaluated by
actual scientists.
“The next leap forward lies in intelligent automation:
the development of multi-agent AI systems that not only process and predict but
also have the technical abilities to emulate perception, reasoning, planning,
and collaboration. These systems have the potential to work alongside human
experts, navigating complex scientific questions, integrating diverse data
sources, generating hypotheses, and designing adaptive experiments. Embracing
this paradigm would allow soil science to harness AI’s full potential—not to
replace researchers but to empower them in addressing the global challenges of
soil security, climate resilience, and sustainable land use.”
The authors give four types
of multi-agent systems they believe can transform soil science research.
1) Integrated
soil digital twin agents – the digital twin can use
algorithms to detect and identify variables that affect nutrient depletion,
water stress, compaction, or erosion hotspots, for instance. The system could
model land management options or soil system responses to climate extremes. It
could generate reports and maps.
2) Soil
microbiome analysis agents – this would require
metagenomics or meta-transcriptomics to evaluate microbial species and be
integrated with detailed environmental parameters such as soil chemistry,
mineralogy, moisture, temperature, or vegetation types. Soil-microbial and
plant-microbial interactions can be unraveled. It could evaluate the effects of
soil amendments or herbicides. It could help in understanding the functions of
the many proteins produced by soil microbes. For instance, it could evaluate
and test nutrient availability under drought conditions. Continent-wide
microbial distributions can be determined and used to evaluate
soil-plant-climate interactions.
3) Climate
change impact and adaptation agents – a system such as this
could integrate climate models with soil process models. Organic carbon stocks
and nutrient and moisture availability could be evaluated. It could evaluate
cropping systems or water conservation techniques. It could evaluate the soil’s
ability to sequester carbon. It could design soil sampling strategies and
optimize fertilization rates. These multi-agent systems can handle complexity
and automate workflows. They warn here that AI cannot replace scientists,
especially when conditions arise that are outside the specificities of their
training.
4) Hypothesis
generation and experimental design agents – here
they note that multi-agentic systems can automate literature analysis and
synthesis and “identify knowledge gaps and propose novel, testable
hypotheses about soil processes.” They could also design experiments to
test those hypotheses.
“Agents should be able to communicate their reasoning,
present findings clearly, and solicit and incorporate expert feedback.”
The
authors next present a case study of hypothesis generation in soil carbon
saturation research. It is mainly a feasibility study for hypothesis
generation. Knowing soil organic carbon (SOC) saturation is important in determining
the capacity of a soil to retain and sequester carbon and for assessing soil
health and fertility. They explain below the debates about carbon saturation
and why testable hypotheses are needed to answer relevant
questions.
“A central concept of organic carbon persistence is soil carbon saturation, particularly regarding mineral-associated organic carbon (MAOC). This concept suggests that mineral surfaces, especially within clay and silt fractions, have a finite capacity to form organo-mineral complexes due to limited reactive surface area. However, its operational relevance is under debate. Empirical studies by Begill et al., Heinemann et al., and Poeplau et al. challenge the idea of a clear saturation plateau, showing near-linear increases in MAOC even under high carbon inputs, implying that biological and environmental constraints may often limit SOC accrual before theoretical saturation is reached. Conversely, Cotrufo et al. maintain the physical validity of the saturation concept, arguing that observed inconsistencies stem from methodological artefacts, such as the misclassification of particulate organic matter (POC) as MAOC, and emphasize the need to identify real-world factors that prevent soils from reaching their saturation potential.”
They utilized Manus AI as
their multitask agent. It has Planner, Execution, and Verification agents that
work collaboratively. The AI system was prompted to scan recent scientific
literature both in and out of the field of soil science to generate contrasting
hypotheses, simulate peer review, and rank them in terms of novelty,
feasibility, plausibility, and scalability.
“In summary, the simulated multi-agent reviews found the
hypotheses to be clear, coherent, and diverse, empirically varied but
scientifically valuable, with several requiring advanced methods for testing,
and well-differentiated with no major redundancy, supporting a broad and
balanced research agenda.”
The hypotheses that are more
difficult to test and those with less scalability across soil types tended to
be ranked lower.
“Table 5 summarizes the conceptual understanding of soil
MAOC saturation based on Georgiou et al. (58), comparing it with the
AI-generated hypotheses.”
Below is a section of the
summary:
“AI systems can help researchers gather information and
rapidly generate hypotheses that are grounded in empirical evidence and
theoretical frameworks. This could reduce the time and effort typically
required to transition from exploratory reading to focused scientific inquiry.”
“Equally important is the enhanced rigor introduced
through simulated peer review. This needs to be further improved through a
step-by-step sequence where researchers evaluate the basis of the hypotheses.
This step requires human input and interaction. For example, soil scientists
stated that they would like to see a more fundamental understanding of
interactions between clay minerals and OM, as most studies use only clay
content to determine or evaluate carbon stock. However, it remains uncertain
whether the hypotheses generated by AI are entirely novel or simply
well-synthesized from existing knowledge. The “black box” nature of the AI
system also does not allow for probing of where the knowledge was derived from.”
In a section about challenges
and future directions, the authors note that data availability, scarcity,
fragmentation, and heterogeneity due to variable sources, as well as lack of
standardization have been identified as challenges.
“Developing and implementing explainable AI (XAI)
methods tailored to soil science applications is essential. Current XAI for
LLMs and generative models mainly provide transparency and provenance rather
than explanations of internal reasoning. While this can improve trust and
usability, it does not offer causal insight into decision processes,
reinforcing the need for human oversight and uncertainty-aware deployment.”
“Predictive uncertainty in multi-agent workflows remains
insufficiently explored. Sensitivity to prompts, stochastic planning, variable
evidence quality, and measurement errors can propagate through agent chains,
increasing the risk of spurious correlations in a data-poor and
methodologically heterogeneous domain such as soil science.”
They also point to
difficulties in integrating AI systems with physical processes such as
automated sample collection, managing sensor networks, and integrating with lab
equipment. Computational costs are another concern.
The quotes below explore the
human-AI interface amid AI’s current constraints.
“Agents must be more than pattern recognizers; they need
a degree of scientific understanding.”
“Creativity as an essential component of the (human)
scientific process should also be nurtured, especially since this remains a
limitation of multi-agent AI systems. AI models are fundamentally dependent on
the data they are trained on. While these agents can autonomously search for
and retrieve literature, their effectiveness is shaped by how they access,
filter, and prioritize sources. This data-dependency constrains their ability
to think “outside the box”, a quality often crucial for scientific breakthroughs.
Although AI excels at recombining and synthesizing existing ideas, its capacity
for true creativity and the generation of genuinely novel concepts remains
uncertain. One potential strength, however, lies in AI’s ability to explore
knowledge across disciplines that are rarely considered by soil scientists,
thereby expanding the conceptual search space. The central question remains:
can AI genuinely create or does it merely reconstruct and reconfigure what is
already known?”
The quotes below are from the
conclusion and continue the assessment of AI capabilities.
“These intelligent collaborators can dramatically
accelerate the pace of discovery, transforming vast datasets into meaningful
insights. By exploring multiple solution pathways and building on existing
knowledge, AI can expand the boundaries of scientific exploration while still
relying on human judgment to interpret, refine, and apply the findings.”
“While AI can emulate aspects of expert reasoning, it
cannot replace the contextual judgment, creativity, and critical interpretation
that human scientists bring to the research process. As such, AI should be
viewed as an augmentative tool, enhancing but not substituting human-led
scientific inquiry.”
References:
Enhancing
soil science research with multi-agent artificial intelligence systems. Budiman
Minasny. Alex McBratney, José A.M. Demattê, Mercedes Román Dobarco, and Pete
Smith. Frontiers in Science. Vol 4. May 20, 2026. Frontiers | Enhancing soil science
research with multi-agent artificial intelligence systems
