Research with AI Agents: How Agentic Systems Are Changing Scientific Work

Agentic AI is changing research

• Agentic systems plan subtasks and use external resources; running generated code for complex calculations can reduce hallucinations from model-internal computation.External resources include web searches and database queries [1]; the computation comparison is described in [2].Source: Agentic AI is changing research · p. 2
• Identifying far-reaching research questions requires deep field context and judgments about both their difficulty and potential benefit.The authors argue that these judgments still have to be made by humans.Source: Agentic AI is changing research · p. 2

The researchers’ role is shifting

• At least 13 percent of biomedical abstracts published in 2024 showed statistically conspicuous word frequencies typical of AI [5].The authors say the actual share probably rose while becoming less obvious.Source: The researchers’ role is shifting · p. 2
• Search makes invented references rarer [6], but roughly a quarter to a half of citations from current web-search models misattribute claims to sources that do not support them [7].Source: The researchers’ role is shifting · p. 2
• In pathology, SPARK generated tissue-parameter hypotheses and executable code, but its search space was limited to cell types and compartments supplied through pathologist annotations [8].Source: The researchers’ role is shifting · p. 2
• The shift mainly concerns digital work, moving researchers from execution to steering and oversight; laboratory work is only partially automated [9].Source: The researchers’ role is shifting · p. 2

Responsibility remains with humans

• Researchers must verify AI outputs and retain responsibility for results, including validating high-risk clinical outputs before use.The authors identify human checking as the bottleneck because it is both time-consuming and demanding.Source: Responsibility remains with humans · pp. 2–3

Efficiency gains through automatable verification

• Efficiency gains are greatest when expected behavior can be checked automatically, shifting oversight toward defining verification criteria.For this article, agents searched and summarized literature with supporting quotations, but the authors selected and read cited works; occasional incorrect or one-sided assessments were corrected.Source: Efficiency gains through automatable… · p. 3, The role of agentic AI in writing this… · pp. 3–4
• Tool-level validation suits fixed tasks with representative data, but research-question agents receive a different task each call, making benchmark averages insufficient.The article describes validation of individual fixed-task modules as one possible approach.Source: Efficiency gains through automatable… · p. 3
• Individual modules can be validated, but the agent’s choices of modules and inputs still require human review; named-source steps make checks more targeted.The oncology decision-support example links intermediate steps to named sources rather than requiring reconstruction of the entire reasoning process.Source: Efficiency gains through automatable… · p. 3

Complex systems elude automated veri fication

• Dynamic agents are assessed mainly empirically and by average performance; reasoning traces can be too long to review, and models may omit bias from their stated reasoning.Research cited in the article finds that LLMs can follow societal biases without documenting them in their chains of thought.Source: Complex systems elude automated veri… · p. 4

What this means for research in pathol ogy

• Incorrect citations in AI-generated text and rising submission and proposal volumes strain review capacity, while ethical AI use could screen low-novelty proposals before submission.The proposed filter compares ideas with literature and may leave fewer, more thoroughly developed proposals for funders to assess.Source: What this means for research in pathol… · p. 4
• Pathology researchers increasingly choose analyses and assess findings, while well-annotated, balanced multicenter datasets increasingly determine the information value of results.SPARK operated within a search space annotated beforehand by pathologists.Source: What this means for research in pathol… · pp. 4–5
• Researchers need to preserve analytical skills and share experience while deciding what to delegate and how to review; problem significance remains a matter of judgment.Predefined criteria cannot determine a research question’s significance for patients, the field, and society.Source: What this means for research in pathol… · p. 5

• The authors declare no competing interests.Source: Compliance with ethical guidelines · p. 6
• Agentic language-model systems supported literature research, checking source support, wording, and critical review; the authors read and checked all cited work and statements.They also state that responsibility for the content remains entirely theirs.Source: Compliance with ethical guidelines · p. 6
• The article reports no studies involving human participants or animals.Source: Compliance with ethical guidelines · p. 6

• Digital research work shifts from execution toward steering and review as agentic systems take on literature search, data analysis, and programming.Source: Wie agentische Systeme… · p. 7
• The greatest efficiency gains arise when expected behavior can be formalized in advance and tested automatically.Systems built from validated modules with traceable combinations can make validation easier.Source: Wie agentische Systeme… · p. 7
• Researchers remain responsible for results and decide what to delegate and how to review outputs.AI assistance alone says nothing about the value of a text, result, or product.Source: Wie agentische Systeme… · p. 7
• The authors recommend institutionalizing exchanges about experience and best practices in using agentic systems.Source: Wie agentische Systeme… · p. 7

Agentische KI verändert die Forschung

• Agents plan subtasks and use external resources [1], while identifying consequential research questions still depends on contextual judgment that delegation may erode [4].Source: Agentische KI verändert die Forschung · p. 8

Die Aufgabe der Forschenden verschiebt sich

• Roughly one quarter to half of citations from current web-search models may not support the specific claims they accompany [7].Search makes fabricated references less common [6], but source-claim mismatches remain.Source: Die Aufgabe der Forschenden verschiebt… · p. 8
• SPARK [8] generated hypotheses and executable code for predicting microsatellite status, but only within a search space prepared through pathologist annotations.The shift chiefly affects digital processes; laboratory tasks are only partly automatable, and researchers move from execution to steering and oversight.Source: Die Aufgabe der Forschenden verschiebt… · p. 8

Die Verantwortung bleibt beim Men schen

• Because output quality still requires demanding human review—and automation complacency can encourage acceptance—researchers remain accountable for validating results.Literature-search summaries can speed selection, but do not replace reading selected papers.Source: Die Verantwortung bleibt beim Men schen · p. 9

Effizienzgewinn durch automatisierbare Prüfung

• Efficiency gains are greatest when code behavior can be tested automatically; fixed-purpose components are easier to validate than agentic combinations.Representative validation data and a well-defined task are prerequisites; validation of agents handling changing research questions remains open beyond benchmark averages.Source: Effizienzgewinn durch automatisierbare… · p. 9

Komplexe Systeme entziehen sich der automatisierten Prüfung

• Dynamic agents resist audit: reasoning traces can exceed human review, and model explanations may omit biases.Current accounts also question language models’ capacity for causal arguments beyond simple formal inference; no operationalizable test for causal understanding is recognized [17].Source: Komplexe Systeme entziehen sich der… · p. 10

Was das für die forschende Pathologie heißt

• AI-written submissions strain peer review and grant evaluation; AI may help filter proposals against literature, but AI-based evaluation can threaten creativity and human goal-setting.The authors also report that AI-generated applications were closer in content to previously funded ideas.Source: Was das für die forschende Pathologie… · pp. 10–11
• Pathology research increasingly depends on curated, balanced datasets and retained analytical skills, while choosing socially important research questions remains human judgment.Institutions should preserve researchers’ ability to design analyses and support exchange of experience and best practices.Source: Was das für die forschende Pathologie… · p. 11

• The authors report no competing interests.Source: Einhaltung ethischer Richtlinien · p. 12
• Agentic language-model systems supported literature research, source checking, wording, and critical review; the authors read cited works and checked claims, retaining full responsibility for the content.Figure 1 was generated with AI and is a schematic illustration without underlying data.Source: Einhaltung ethischer Richtlinien · p. 12
• The article reports no studies involving humans or animals.Source: Einhaltung ethischer Richtlinien · p. 12

Danksagung

• Six people are thanked for discussions that shaped the article’s position and argument.Daniel Budelmann, Stefan Heldmann, Nils Papenberg, Jan-Philip Redlich, Raphael Schäfer, and Nick Weiss.Source: Danksagung · p. 12
• Gabriele Lotz, Ole Schwen, and Till Nicke are thanked for critically reviewing the manuscript.Source: Danksagung · p. 12
• The project was funded by Germany’s Federal Ministry of Education and Research under grant 03VP13371.The authors state that they are responsible for the publication’s content.Source: Danksagung · p. 12

Literaturverzeichnis / References

• A scoping review addresses AI agents in healthcare research [1].Source: Literaturverzeichnis / References · p. 12
• The bibliography includes PAL, a program-aided language-model paper [2].Source: Literaturverzeichnis / References · p. 12
• Sources cover DFG guidance on AI in peer review [3] and Tao’s discussion of mining fruitful open problems [4].Source: Literaturverzeichnis / References · p. 12
• A study examines excess vocabulary in LLM-assisted biomedical writing [5].Source: Literaturverzeichnis / References · p. 12
• Two papers address reference hallucinations and source attribution in research agents [6, 7].Source: Literaturverzeichnis / References · p. 12
• Pathology references cover autonomous cancer discovery [8] and laboratory automation [9].Source: Literaturverzeichnis / References · p. 12