Reframing Discovery Trade-Off in Plant Genomics Through Autonomous Agents

1. The Trade-Off Frontier in Plant Genomics

Discovery in plant genomics is rarely reducible to a single model invocation or database query, but a sequence of interpretive decisions about prior evidence, genome annotations, statistical assumptions, cross-species comparability, and validation priorities (Dominguez Del Angel et al., 2018). The trade-off frontier arises because these decisions are cumulative and path dependent: each requires time and expertise, and errors or omissions at one stage can propagate through the workflow and reshape the biological conclusion (Salzberg, 2019).

1.1. Three Operational Dimensions of the Frontier

We describe this constraint space along three operational axes: literature depth, analytical breadth, and time efficiency (Figure 1). These axes are not intended to exhaust all sources of variation, but to capture recurring questions in real projects: how much relevant knowledge was incorporated, how broadly the analytical landscape was explored, and whether the workflow produced a result quickly enough to guide the next experiment. In practice, each axis is evaluated through operational proxies rather than treated as a single scalar metric.

Literature depth refers to the extent to which prior knowledge is incorporated into analysis. In candidate-gene studies, this includes recovery of known homologues and prior functional evidence, together with trait associations, expression evidence, and contradictory findings (Kaňovská et al., 2024). Operational proxies include recall of known gene–trait associations within a fixed time budget and traceability of claims to cited evidence. Downstream analyses often draw on only a selective subset of available knowledge because gene names change, crop-specific databases remain incomplete, and relevant evidence is scattered across species (Olson et al., 2026).

Analytical breadth captures how much of the computational landscape a workflow explores. A GWAS workflow, for example, may span data preprocessing, association testing and peak definition, candidate annotation, and downstream evidence integration (Xu et al., 2025a). Broader workflows may additionally consult pan-genome summaries, structural variants, or gene presence–absence matrices. Each added component can improve biological context, but it also introduces dependencies, file-format conversions, and new failure modes.

Time efficiency is the person-time required to move from a biological question to an actionable computational result. Runtime is only one component: Researchers also spend substantial time on data preparation, tool configuration, error diagnosis, parameter tuning, and interpretation of intermediate outputs (Wratten et al., 2021), which for many interpretive workflows dominate the budget (Michael, 2026).

These axes are deliberately operational rather than purely numerical. Literature depth is not simply the number of papers cited, analytical breadth is not the number of tools invoked, and time efficiency is not compute time alone. The value of the frontier trade-off is that it asks whether a workflow improves these dimensions jointly, not at a single convenient point.

1.2. Why the Constraint Persists

Cross-species inference makes the coupling among these dimensions concrete. Functional evidence from rice may or may not be relevant to maize, wheat, potato, or soybean. Its relevance depends on orthology confidence, regulatory conservation, and trait context (Gabaldón and Koonin, 2013). Assessing these conditions may require additional literature review, database reconciliation, and computational analysis. Manual workflows can address these steps carefully, but rarely at scale or with complete reproducibility.

The constraint persists because the burden of integration is structurally distributed. Plant knowledge is fragmented across genome versions, databases, laboratory traditions, and publication types (Heuermann et al., 2025). Computational workflows span programming languages, operating systems, file formats, and parameter conventions (Reiter et al., 2021). Experimental decisions are in turn distributed across molecular biologists, bioinformaticians, and field teams (Robinson et al., 2024). This overhead is often invisible in publications, but it determines which analyses are attempted, repeated, or abandoned.

This distribution makes results path-dependent: small analytical choices can lead to different biological conclusions (Stark et al., 2019). Plant-genomics evidence is also rarely decisive: a candidate gene is usually supported by several incomplete signals rather than one definitive assay, and teams can reasonably weigh those signals differently, especially for environmentally sensitive traits or regulatory variants. Reproducing such a result means recording how the evidence was assembled, not merely rerunning the final command.

1.3. What Would Count as a Shift

Conversational language models can help researchers understand methods, summarise papers, and draft analysis plans, although reliable end-to-end execution of complex bioinformatics workflows remains challenging (Nikita et al., 2025). Specialised methods, including GWAS models, genomic prediction frameworks, and network analyses, improve performance along individual dimensions of the workflow. Yet researchers still need to coordinate these methods, reconcile their outputs, and carefully document each decision: the challenge lies not in a lack of software, but in the significant cost of coordinating it effectively within reviewable workflows.

Shifting the frontier would require capabilities that act on the three axes together: retrieval and tracing of fragmented evidence across species and data types; coordination of diverse tools and data modalities within a reproducible process; and repeated computational cycles quickly enough to support hypothesis refinement rather than one-pass analysis. The relevant unit of evaluation is therefore an end-to-end workflow, not a single prompt, database query, or tool call (Dionizije et al., 2026).

Agents do not remove trade-offs, but they may change their scale by externalising routine cognitive load and making repeated analysis easier to audit (Lu et al., 2026). The key question is whether agentic systems can expand the feasible region of plant-genomics discovery while humans remain responsible for interpretation, experimental design, and high-cost biological decisions. Answering it requires a capability framework organised around the scientific work a system performs, rather than the architecture it uses.

2. A Capability Hierarchy for Autonomous Agents in Plant Genomics

A useful hierarchy should classify what a system does for plant-science discovery, not which model architecture it uses. The relevant questions are functional: whether the system recovers overlooked evidence, makes a multi-tool workflow reproducible, or revises a candidate ranking after new computational evidence appears. The L1–L4 hierarchy below is organised around these scientific tasks, complementing general research-agent taxonomies that classify systems by autonomy, responsibility allocation and workflow closure (Tie et al., 2026), while asking which plant-genomics bottleneck each level can demonstrably relieve.

2.1. Levels Defined by Scientific Work

A system belongs at a given level because of what it demonstrably accomplishes in a plant-science task. L1–L3 describe capabilities that can in principle be tested now, whereas L4 is retained as a long-term upper bound. Table 1 defines L1–L4 by minimum qualifying criteria and boundary conditions, summarises current evidence status, and specifies the human role at each level.

Unaugmented conversational language models provide a practical baseline but not capability level, because they do not reliably retrieve traceable evidence, execute workflows, or preserve state across iterative analysis (Chelli et al., 2024; Zhou et al., 2025). The hierarchy (Figure 2) begins at L1, where external knowledge sources are retrieved and cited; moves to L2, where external tools are executed and logged; and reaches L3 only when hypothesis, computation, and refinement form a closed loop. L4 is reserved for self-evolving systems that improve across repeated projects, for which plant-specific evidence remains absent.

The hierarchy also separates capability from evidence status. An architecturally sophisticated system provides weak plant genomics evidence if untested on plant tasks, whereas a narrower retrieval system can support a strong L1 claim if it performs well on crop-specific questions with traceable citations. This distinction matters where many precedents come from biomedicine, chemistry, protein design, or single-cell analysis (Asai et al., 2026; Dip et al., 2026; Qi et al., 2026). Each level is therefore defined by falsifiable criteria. L3, for example, requires at least three autonomous hypothesis–test–refinement cycles on a defined task, which separates iterative discovery from single-step automation.

2.2. L1: Knowledge Integration

L1 systems connect distributed biological knowledge into traceable evidence (Asai et al., 2026; Yu et al., 2026). In plant science, this matters because relevant knowledge is distributed across crop-specific databases, genome versions, gene identifiers, and experimental contexts. A single biological signal may surface as a cloned gene in one crop, a QTL candidate in another, a stress-responsive transcript in an expression dataset, or a weakly annotated orthologue in a related species. Current plant-focused systems illustrate different L1 strategies, which primarily differ along two dimensions: knowledge representation and update mode (Table S1).

Regarding knowledge representation, systems can be divided into graph-centred and corpus-centred approaches. Graph-centred systems, represented by SeedLLM.Rice and PlantScience.ai, store gene–trait–pathway relationships as explicit nodes and edges, making them suited to tests of relational recovery and citation-grounded evidence assembly (Yang et al., 2025b; Yu et al., 2026). They should therefore be judged by relational accuracy as well as explanation fluency. Conversely, corpus-centred systems retrieve and synthesise text directly from the literature, preserving the context of original claims but often leaving relationships implicit. PlantGPT illustrates this pattern in Arabidopsis functional genomics through retrieval-augmented generation (Zhang et al., 2025b), while OpenScholar provides a cross-domain comparator for retrieval-augmented literature integration and citation evaluation (Asai et al., 2026). Such systems are therefore better assessed by citation precision, coverage and the ability to surface conflicting evidence. Regarding update mode, a curated resource like Potato Knowledge Hub (Li et al., 2026) provides stable and auditable functional-genomics knowledge, whereas a continuously updated system like PlantScience.ai maintains timeliness but requires stronger checks for duplicate identifiers, outdated annotations, and unverified claims (Yu et al., 2026).

The most informative L1 evaluations are therefore gene- and trait-centred rather than based on question-answering demonstrations alone. A system asked to prioritise drought-tolerance candidates should separate direct crop evidence from model-species analogy, distinguish association evidence from functional validation, avoid treating expression correlation as causal proof, and flag uncertainty around gene identifiers or genome versions. These checks are essential, because an apparently precise gene name can mask orthology ambiguity, annotation drift, or a family-level signal that should not be attributed to a single locus (Shi et al., 2023). By surfacing missed orthologues, contradictory phenotypes, or annotation problems, L1 systems can improve evidence gathering before downstream analysis, but do not by themselves constitute autonomous discovery.

2.3. L2: Tool Orchestration

L2 systems move from evidence retrieval to execution by turning tools, code, and intermediate files into repeatable workflows with limited human intervention. This capability addresses the gap between a biological question and a reproducible computational pipeline (Nikita et al., 2025). A typical workflow may involve command-line tools, R packages, Python scripts, genome-specific files, and manual quality checks (Wratten et al., 2021); L2 agents aim to make this scattered activity traceable.

Plant-domain L2 evidence is still early and heterogeneous (Table S1). PhenoAssistant illustrates phenotype-oriented tool orchestration by connecting a language-model interface to curated tools for phenotype extraction, visualisation, and model training (Chen et al., 2026). ABC (Agricultural Breeding Claw) and BOLE extend this pattern toward natural-language orchestration of breeding workflows, though both remain prototypes with limited systematic evaluation (volcanooooooooo, 2026; Wang et al., 2026). GEAIR provides a stronger plant-domain example of coupling genome editing with robot-assisted pollination, but its validation target is breeding automation rather than general genomics workflow orchestration (Xie et al., 2025). Cross-domain systems are useful mainly as mechanism precedents: GeneGPT and BioAgents illustrate database and tool access; AutoBA, DrBioRight 2.0 and SciToolAgent illustrate workflow automation and tool selection; ProteinMCP and BioinfoMCP illustrate protocol-based tool interfaces; STAgent and SpatialAgent add spatial-biology examples of the same orchestration pattern (Lin et al., 2025; Wang et al., 2025a); and transcriptomic or multi-omics systems such as SeqMate and PromptBio show reusable execution templates in adjacent bioinformatics settings (Devam and Atharva, 2024; Ding et al., 2025; Jin et al., 2024; Lei et al., 2024; Liu et al., 2025; Nikita et al., 2025; PromptBio, 2025; Widjaja et al., 2025; Xu et al., 2026; Zhou et al., 2024).

Established workflow management systems (WMS, such as Snakemake, Nextflow/nf-core, Galaxy) provide an important baseline for L2, but they do not decide which tools are appropriate, how to interpret an ambiguous input, or when a biological assumption has been violated (Alam and Roy, 2025; Community, 2024; Ewels et al., 2020; Mölder et al., 2021). An L2 agent operates above this layer: translating a biological question into a workflow and adapting it when inputs are imperfect. In a post-GWAS analysis, for example, a WMS can run a preset workflow for peak delineation, gene extraction, identifier harmonisation, and expression summarisation; an L2 agent should additionally recognise a mismatched genome build, record why a default interval was chosen, and trigger expert review when the inconsistency affects interpretation (Kaňovská et al., 2024). The two are therefore complementary: a well-designed L2 agent should build on WMSs, inheriting reproducible execution while adding accessibility, monitoring, and adaptive control.

2.4. L3: Closed-Loop Computational Discovery

L3 systems perform iterative cycles of hypothesis generation, computational testing, and refinement, in which an agent begins to resemble a computational research collaborator rather than a searchable assistant or workflow runner. We define a minimal criterion as at least three consecutive hypothesis–test–refinement cycles executed without human intervention on a defined plant-science task. Each cycle must articulate a hypothesis, retrieve evidence or run a test, and update a ranking or interpretation; the next cycle must use the previous output as input rather than restart from the original prompt. The threshold is intentionally conservative: it distinguishes closed-loop discovery from single-pass automation, while remaining measurable in a benchmark setting.

To our knowledge, no peer-reviewed plant system yet satisfies this criterion in a general and well-benchmarked way. Aleks is one of the closest plant-domain attempts, reporting autonomous, multi-cycle problem formulation and model refinement, although only within a single exploratory case study (Jin et al., 2025a). Cross-domain examples nevertheless demonstrate that iterative scientific agents are technically possible. The AI Scientist automates idea generation, experiment execution, and analysis in machine-learning research (Lu et al., 2026); Co-Scientist and Robin show multi-agent scientific discovery workflows in biomedical and experimental biology settings (Ghareeb et al., 2026; Gottweis et al., 2026). Chemical, genome-editing, single-cell and general research-agent systems extend the landscape (Table S1), but their feedback loops, validation labels and task structures differ from plant-genomics candidate refinement.

The distinction between L3 and L2 is not the number of steps, but whether outputs from one analytical cycle determine the next test. Even a complex GWAS-to-report workflow remains L2 if it follows a preset sequence and stops at summarisation; it becomes L3 only when the system revises its next test after evidence such as contradictory mutant data, paralogy, or tissue-specific expression undermines the original hypothesis. The output should be a more defensible set of hypotheses for expert review, not experimental confirmation. Establishing causal function still requires wet-lab or field validation, so L3 quality should therefore be judged by whether successive cycles make the reasoning trajectory more robust.

2.5. L4: Self-Evolving Systems

L4 systems would adapt their own capabilities over time, for example by generating new tools, improving workflow templates, or creating specialised sub-agents (Bu et al., 2026). In plant genomics, this might eventually mean a system that learns from repeated gene-discovery projects across crops and updates its evidence-gathering strategy. Recent BioMedAgent and STELLA illustrate biomedical precedents for adaptive agent behaviour, while SkillEvolver and EmbodiSkill offer more general mechanisms for skill evolution; however, none has yet been evaluated on plant-genomics tasks (Table S1) (Bu et al., 2026; Jin et al., 2025b; Ju et al., 2026; Zhang et al., 2026). L4 is therefore included as a conceptual upper bound rather than a near-term objective. Practical progress should be judged first at L1 and L2, and then through carefully benchmarked L3 demonstrations.

3. Technical Substrate for Auditable Agentic Workflows

3.1. From Plant-Science Pain Points to Agent Paradigms

Agentic systems are easiest to evaluate when they are tied to specific plant-science bottlenecks. Three recur in plant genomics, each calling for a different capability level (Table 1): scattered evidence (knowledge dispersed across literature, databases, genome versions and identifiers) calls for evidence reconciliation with traceable citations (L1); analyst-dependent workflows call for logged, reproducible tool orchestration (L2); and slow iterative refinement calls for closed-loop cycles in which each test updates the next (L3) (Berger and Yu, 2023; Heuermann et al., 2025; Olson et al., 2026). This mapping keeps the term “agent” from becoming a generic label: each bottleneck maps to a different level, a system should be evaluated at the level its evidence supports, and whatever the level, the workflow must remain auditable (Table 2).

3.2. From Tool Use to Auditable Execution

Agentic workflows become auditable only when execution is organised rather than improvised. In bioinformatics, this means separating three kinds of action: calling established tools, reusing tested workflows, and generating code only for local tasks that are not already covered by a workflow (Table 2) (Nikita et al., 2025). The purpose of this separation is to keep the path from biological instruction to computational output visible.

Tool calling allows an agent to invoke configured programs without requiring the user to manually switch environments. For a plant breeder or experimental biologist, the key point is that a multi-step analysis can be executed, logged, and repeated from a biological instruction. Interoperability efforts such as the Model Context Protocol (MCP), described as an emerging standard for connecting language models to external tools and resources, can reduce fragmentation in tool access (Hou et al., 2025). Reusable agent skills and workflow templates represent a related but less standardised implementation strategy.

Reusable workflows provide the stable substrate (Wratten et al., 2021). Agent skills can encapsulate standard analytical protocols, and established WMS can supply auditable execution contexts for multi-step bioinformatics (Yang et al., 2025c). In this framing, the agents can call, inspect, or adapt the workflow infrastructure (Alam and Roy, 2025), which reduces dependence on one analyst’s habits and make local expertise reusable.

Code generation should be the exception. It is useful for local transformations, such as identifier conversion, table reshaping, or customised visualisation, where no predefined workflow exists (Tang et al., 2024), but it is also the easiest way to embed unreviewed assumptions into a workflow (Zhang et al., 2024). Generated code should therefore remain sandboxed, logged, and reviewable; an agent that silently generates code or overwrites intermediate files cannot be considered auditable, even if its final report appears plausible.

3.3. Scope and Limits of Cross-Domain Evidence

Cross-domain studies show that many components of agentic systems are technically feasible (Dip et al., 2026; Qi et al., 2026). Yet autonomy is domain-conditioned, depending on reliability, provenance, validation speed and accountability (Tie et al., 2026). The question for plant genomics is which patterns from adjacent biological fields transfer to crop biology, and which depend on evidence structures that differ from those of plant science.

The transferable lessons are mainly infrastructural: standardised tool interfaces, provenance and audit logging, retrieval-augmented reasoning, and planner-executor-verifier decompositions. Recent systems and benchmarks suggest that retrieval-augmented reasoning, evidence grounding and process-level evaluation are important design features for biological agents (Asai et al., 2026; Dionizije et al., 2026; Mitchener et al., 2025). Biomedical and bioinformatics systems have connected language models to domain-specific databases, tool registries, code execution, workflow templates, and report generation (Dip et al., 2026; Huang et al., 2025; Jin et al., 2025b). Single-cell, spatial-biology and virtual-cell agents such as CellAgent, SpatialAgent and CellForge add a methodological precedent: multi-step biological analyses can be benchmarked against published analytical decisions rather than only demonstrated on curated examples (Tang et al., 2025; Wang et al., 2025a; Xiao et al., 2024).

Other patterns transfer poorly. Protein-design, drug-discovery, therapeutic-reasoning and clinical-agent studies, such as ProtAgents, DrugAgent, PharmAgents and MedAgents, show that agentic architectures can be adapted to specialised biomedical decision contexts (Table S1) (Gao et al., 2025a; Ghafarollahi and Buehler, 2024; Liu et al., 2024b; Tang et al., 2023). Their feedback loops, safety constraints and validation labels differ sharply from crop functional genomics. Single-cell biology has denser benchmark labels than crop functional genomics; protein engineering and some biomedical design tasks allow fast in silico or in vitro feedback, whereas plant validation often requires seasonal field cycles; and closed-loop wet-lab paradigms from synthetic chemistry do not transfer straightforwardly for crop breeding (Tom et al., 2024; Xu et al., 2025b; Zhang et al., 2025a). Cross-domain evidence therefore supports technical feasibility for L1 and L2, and plausibility for L3 computational pre-screening, but not closed-loop experimental discovery in plant systems. Plant-specific benchmarks are needed to determine which patterns actually improve discovery rather than merely reproduce a familiar agent architecture.

3.4. Reliability and Boundary Conditions

Reliability matters when agent outputs influence experimental design or breeding decisions. Useful safeguards include database-grounded self-verification, multi-agent or adversarial review to identify unstable conclusions, human-in-the-loop escalation for high-cost or irreversible decisions, and provenance-aware logging of workflow composition (Table 2) (Ghareeb et al., 2026; Wang et al., 2025b). GeneAgent offers a concrete pattern, in which gene-set claims are checked against domain databases before being reported (Wang et al., 2025b). Comparable safeguards are especially important in plant genomics when workflows cross genome builds, gene-model releases, or coordinate systems. Reliability should be evaluated at the workflow level, including reproducibility across runs, supervision required, and provenance completeness.

Some constraints remain even when agentic workflows become more capable. Non-model crops lack validation data; genomic signals rarely determine phenotype on their own; and genotype-by-environment interactions with field heterogeneity complicate causal interpretation (Xu et al., 2025b). Commercial breeding programmes also face privacy, access-control and data-governance requirements that make unrestricted data sharing or deployment difficult (Gogna et al., 2025). These limitations define where agents belong: evidence integration, reproducible computation, and prioritisation, not replacement of biological validation or expert judgement.

3.5. From Case Studies to Benchmarks

Individual case studies show that agentic workflows can be assembled, but they do not by themselves create field-level evidence (Mitchener et al., 2025). Evidence accumulates only when tasks, inputs, scoring rules, and baselines are comparable across systems. Recent benchmarks move evaluation beyond fact recall toward workflow execution, code generation, dataset exploration, marker-evidence attribution and robustness under perturbation (Table S1), where PlantMarkerBench is especially relevant as it scores evidence-grounded marker reasoning across plant species (Acharjee Dip et al., 2026). These benchmarks also expose current limits: an agent may still produce a plausible plan while failing at data handling, step-level reasoning, evidence grounding or reproducibility (Dionizije et al., 2026).

Plant science needs analogous benchmark tasks with public inputs, defined outputs, transparent scoring and independent baselines (Mitchener et al., 2025). We therefore propose PlantAgentBench (Box 1, Figure 3) as a plant-specific framework that samples the three workflow bottlenecks: evidence integration, reproducible multi-tool execution, and iterative computational refinement. The numerical scales and scenarios in Box 1 are design parameters rather than fixed requirements, and the task family sets should not be interpreted as covering the full diversity of plant-genomics discovery. Future versions can incorporate pan-genome-aware interpretation, structural variation, phenomics, and genomic-prediction diagnosis when suitable datasets, scoring rules, and expert standards are available. At minimum, PlantAgentBench should combine public and hidden tasks, independent baselines, standardized inputs, transparent scoring and expert adjudication that records disagreement.

Box 1. PlantAgentBench task families for evaluating autonomous plant-genomics agents. PlantAgentBench maps the L1–L3 hierarchy onto three recurring plant-genomics bottlenecks: literature-based gene prioritisation, statistical-genomics workflow execution and iterative computational hypothesis refinement. For each task family, the table defines representative inputs, expected outputs, reference baselines, stressors and task-specific metrics. The scenarios are design examples rather than fixed requirements; benchmark-design conventions are discussed in Section 3.5.

Task Family (Target Level)	Inputs/Scenario	Outputs	Baseline	Stressors	Metrics
Literature-based gene prioritisation (L1)	Candidate-gene list, fixed literature/database snapshot, identifier metadata (e.g., rice drought tolerance)	Ranked genes with evidence summaries and citations	Expert consensus, known validated genes	Synonyms, orthology ambiguity, contradictory evidence, outdated annotations	Recall@5/10; MAP; citation precision; provenance completeness; unsupported-claim rate
Statistical-genomics workflow execution (L2)	Genotype matrix, phenotype table, genome annotation, expected formats (e.g., maize GWAS follow-up)	QC summary; Manhattan/QQ plots; significant SNPs; effect sizes; logs	Expert-curated reference workflow; WMS-only; published outputs	Missing metadata, genome-build mismatch, file-format variation, parameter ambiguity	Automated steps; reproducibility; effect-size correlation; SNP-list agreement; intervention count
Iterative computational hypothesis refinement (L3)	Initial association signals, multi-omics + causal evidence, stopping rules (crop-trait refinement)	Final ranked genes; evidence trail; iteration log with decision trace	Expert panel with disagreement recorded; validated functional evidence where available	Conflicting mutant data, paralogy, tissue-specific expression, partial evidence	Autonomous cycles; evidence-type coverage; NDCG@10; runtime per cycle; intervention count

PlantAgentBench should evaluate process as well as output. Its core requirement is not merely a correct gene ranking, but an auditable reasoning path: a correct gene ranking without citations is incomplete; a Manhattan plot without recorded software versions and parameters is not reproducible; a final candidate list without an iteration log cannot distinguish iterative refinement from a longer one-pass report (Acharjee Dip et al., 2026; Huang et al., 2025; Qu et al., 2026b). These requirements match how plant scientists decide whether a computational result can guide experimental work.

Benchmark design must avoid two traps. Overly clean tasks reward systems that perform well on curated examples but fail on real crop data; overly open tasks are difficult to score and may reward verbose reasoning over accurate prioritisation. PlantAgentBench therefore combines controlled tasks with known or expert-curated outputs and messier tasks judged by expert panel.

Governance is also part of benchmark design. Public tasks are needed for reproducibility and method development, but hidden or periodically refreshed tasks are needed to limit benchmark overfitting. Gold standards should record expert disagreement rather than hide it, since candidate-gene prioritisation often involves partial evidence. Reporting should be stratified by crop, trait type, and task level: a single aggregate score would obscure systems that perform well on literature retrieval but fail at workflow execution, or systems that execute workflows but remain weak at biological interpretation.

4. Challenges, Roadmap, and Call to Action

4.1. Challenges to Deployment

The main barriers are not peripheral. Long validation cycles magnify reliability problems, crop diversity magnifies data-standardisation problems, and the commercial value of breeding datasets magnifies governance issues. These risks extend existing weaknesses in plant-genomics workflows, including irreproducible pipelines, inconsistent gene-trait mappings, and undocumented analytical choices (Nikita et al., 2025). Agents can amplify these existing problems or help mitigate them, depending on the constraints under which they operate (Table 3).

Adoption will depend on risks and incentives. Literature triage is a low-risk entry point, whereas candidate nomination, breeding decisions, and other high-cost choices require stronger validation and explicit human approval. Breeding programmes may also value local deployment and inspectable decision traces more than marginal gains in model capability. A staged adoption path can therefore accumulate evidence without allowing immature systems to influence expensive biological decisions prematurely.

4.2. A Staged Roadmap

The roadmap should be read as a set of proposed evaluation milestones rather than predictions (Figure 3).

Near-term milestone (2025–2026): L1 and L2 deployment. Crop-specific knowledge systems and tool-orchestration agents should be evaluated in routine gene-discovery and phenotyping workflows. Useful milestones include independent demonstrations of improved literature-search efficiency, citation traceability, workflow reproducibility, or reduced intervention count.

Mid-term milestone (2027–2028): L3-like computational refinement. The key milestone is a peer-reviewed demonstration of at least three autonomous computational refinement cycles in a plant-genomics task, with documented improvement across cycles and limited human intervention. The output should be a candidate ranking with an evidence trail, not a claim of experimental proof.

Long-term milestone (2029–2030): adaptive agent ecosystems. Longer-term progress may involve systems that improve reusable workflows, generate analysis modules, or learn from accumulated laboratory experience. Adoption should be judged by independent reuse and reproducibility rather than by novelty alone.

These milestones also remain revisable. Weak citation precision should return the priority to curation and evidence representation, and automation that misses common data mismatches should not be mistaken for workflow reliability. Failure to meet the L3 criterion on public plant-genomics tasks should lead to revision of the hierarchy rather than relabelling weaker evidence as L3.

4.3. Cost, Privacy, and Local Deployment

Although model access is becoming easier, peer-reviewed quantification of biological-agent deployment costs remains limited (Qi et al., 2026). For plant-science laboratories, the limiting factor is likely to be engineering integration, data harmonisation, validation infrastructure, and governance rather than model access alone. Local or institution-specific deployment may be especially important, because unpublished germplasm data, field-trial records, and proprietary breeding information cannot always be sent to external services. Smaller specialised models, local knowledge graphs, and private tool registries may therefore be more valuable than general cloud systems when reproducibility and data control are priorities.

4.4. Immediate Priorities and Unresolved Questions

Three priorities follow directly from the framework. First, a benchmark consortium for plant-specific agentic workflows across model crops and underrepresented crops, with PlantAgentBench as a shared community instrument. Second, data standardisation and interoperable infrastructure are equally scientific investments: harmonised gene identifiers, stable annotation mappings, provenance-aware databases, and tool interfaces will determine whether agents can reason across crops reliably. Third, training for human–agent collaboration: the relevant skill is auditing evidence trails, recognising biological implausibility, inspecting workflow provenance, and deciding when an automated result is ready for experimental follow-up.

Several questions remain outside the reach of the framework. The biological limits of cross-species functional transfer are real, genotype-by-environment interactions in field phenotyping are not mere engineering inconveniences, and breeding-data governance cannot be solved by better models alone. Responsibility for agent outputs must also be allocated among the analyst, principal investigator, institution, and system provider. The value of plant-genomics agents will depend on whether they operate transparently and safely within these constraints.

Conclusions

The trade-off frontier in plant genomics, spanning literature depth, analytical breadth, and time efficiency, captures a structural feature of discovery work. Plant scientists increasingly have access to rich datasets and powerful methods, but the work of connecting evidence, executing analyses, and revising hypotheses remains difficult to scale.

The plant-genomics L1–L4 hierarchy links agent capabilities to specific bottlenecks. L1 systems address fragmented knowledge; L2 systems address tool coordination and reproducibility; L3 systems define a testable target for autonomous computational discovery; and L4 remains a long-term conceptual upper bound. The hierarchy is valuable because it prevents vague claims about AI capability by making visible the constraint relieved, the benchmark used, the provenance recorded, the evidence supported, and the level of human intervention required. PlantAgentBench translates this argument into an evaluation agenda, where agents that aspire to be trusted infrastructure must be tested on tasks that resemble real gene-discovery workflows: retrieving evidence, executing analyses, revising candidate rankings, and preserving a decision trace. Such benchmarks would allow the field to tell whether agents truly expand the feasible region of discovery or merely make existing steps more convenient.

Autonomous agents will not replace domain expertise. The more useful question is whether auditable, benchmarked, and domain-grounded agents can extend the reach of expert reasoning. If they can, the next phase of plant genomics may be defined not only by larger datasets or more accurate models, but by a better organised discovery process.