Agentic Scientist: Integration and Research Autonomy

• The Agentic Scientist is an integrated system that combines large language models, multimodal models, retrieval systems, simulators, code interpreters, and more to automate and optimize scientific workflows.
• By leveraging Reasoning and planning, tool integration, memory, multi-agent collaboration, and optimization; this system enables iterative and adaptive research processes across various scientific domains.
• Applications include drug development, chemical research, material sciences, and fluid dynamics, with current systems demonstrating varying degrees of autonomy and procedural capabilities.

The Agentic Scientist is a goal-directed, tool-connected research system that coordinates substantial portions of the scientific lifecycle: problem formulation, literature retrieval, hypothesis generation, experimental or computational design, execution, analysis, validation, revision, and communication. Unlike a conventional language-model assistant, it operates through iterative perception, reasoning, planning, action, and feedback. The term denotes a functional role performed by an integrated system rather than a single model, typically combining LLMs, multimodal models, retrieval systems, scientific databases, simulators, code interpreters, laboratory instruments, verification modules, persistent memory, and human oversight. Current systems demonstrate substantial procedural autonomy in constrained environments, but no cited system reliably realizes a universally autonomous, open-ended scientist (Gridach et al., 12 Mar 2025).

1. Conceptual foundations and autonomy levels

Agentic Science is positioned as a stage in the evolution of AI for Science: the transition from computational tools to active, goal-directed research partners. A mature Agentic Scientist should be able to observe a domain, formulate non-obvious hypotheses, design and execute experiments, analyze evidence, validate conclusions, and iteratively refine its knowledge and strategy with minimal human intervention (Wei et al., 18 Aug 2025).

Scientific agency is distinguished from conversational fluency and isolated tool use. It requires purposeful action in a scientific environment, empirical feedback, memory of prior operations, adaptation over long horizons, and accountability through provenance and verification. The defining characteristic is therefore closed-loop action rather than text generation alone.

A useful autonomy taxonomy comprises four levels:

1. AI as a computational oracle: a static model performs a predefined prediction or generation task. Humans select the problem, formulate hypotheses, choose experiments, interpret results, and determine subsequent actions.
2. AI as an automated research assistant: an agent executes predefined workflows, chains tools, generates code, runs analyses, or controls restricted experimental processes. It possesses task-level autonomy but not goal-level autonomy.
3. AI as an autonomous scientific partner: an agent pursues high-level scientific objectives through the complete discovery loop, changes hypotheses and experimental strategies, and uses evidence to update its beliefs.
4. AI as a generative architect: a prospective level at which an agent invents new instruments, methods, mathematical frameworks, or scientific disciplines.

Most deployed systems remain at Level 2 or occupy restricted portions of Level 3. They may automate a complete workflow within a domain, but generally depend on human-supplied objectives, datasets, code interfaces, computational environments, safety constraints, or validation criteria. The distinction between task-level and goal-level autonomy is consequently central: an agent may decide how to solve a problem without deciding which scientific problem is worth solving.

A position paper challenges the assumption that increasing procedural autonomy necessarily produces scientific autonomy. It distinguishes co-scientists, which extend human judgment on well-specified tasks, from autonomous scientific discoverers, which independently select worthwhile questions, maintain diverse hypotheses, learn from failure, revise objectives, and remain answerable to physical evidence (Bisht et al., 9 May 2026). Another philosophical critique argues that scientific interpretation and authorship cannot be delegated without weakening the human creativity, understanding, and responsibility that make research trustworthy (Hogg, 31 Aug 2026).

2. Capabilities and system architecture

An Agentic Scientist is normally a modular system whose components support five foundational capabilities:

• Reasoning and planning: decomposition of scientific objectives, hypothesis formation, experimental design, tool selection, backtracking, reflection, and adaptive planning. Mechanisms include ReAct-style loops, Tree-of-Thought search, Monte Carlo Tree Search, self-consistency, debate, and reinforcement learning.
• Tool integration: invocation of literature databases, code interpreters, symbolic systems, molecular-property predictors, chemistry toolkits, simulators, molecular docking systems, bioinformatics pipelines, robotic laboratories, and scientific instruments.
• Memory: retention of hypotheses, experimental conditions, negative results, tool traces, causal histories, and reusable procedures. Relevant mechanisms include retrieval-augmented generation, knowledge graphs, episodic memory, skill libraries, and persistent-memory architectures such as Letta and MemGPT.
• Multi-agent collaboration: division of labor among planners, literature researchers, experiment designers, executors, evaluators, critics, analysts, and report writers. Multi-agent systems improve role specialization but introduce communication overhead, inconsistent representations, coordination failure, error propagation, and responsibility ambiguity.
• Optimization and evolution: iterative self-refinement, execution-based debugging, self-rewarding reinforcement learning, red-teaming, population-based search, and improvement of tools, strategies, or agent architectures.

These capabilities enable a dynamic four-stage workflow:

1. Observation and hypothesis generation: the system retrieves literature, databases, prior experiments, and multimodal observations; identifies gaps, contradictions, and unexplained patterns; and ranks candidate hypotheses according to novelty, plausibility, feasibility, and expected utility.
2. Experimental planning and execution: the selected hypothesis is converted into a protocol, simulation plan, computational workflow, or laboratory intervention with controls, parameters, measurements, stopping criteria, and safety constraints.
3. Data and result analysis: the system processes numerical, textual, visual, chemical, biological, or physical evidence and evaluates whether it supports the hypothesis.
4. Synthesis, validation, and evolution: the system performs critique, replication, consistency checking, and uncertainty assessment; accepts, rejects, or modifies the hypothesis; updates memory; and revises future strategy.

The stages may be repeated, skipped, or reordered. A mature system must distinguish execution success from scientific validity: a program can run successfully while manipulating the wrong variable, omitting a baseline, producing an invalid comparison, or supporting an unjustified conclusion.

Infrastructure-oriented work extends this architecture beyond the agent itself. Bohrium and SciMaster describe a stack in which scientific data, software, compute, laboratory systems, workflows, and validation procedures become agent-ready, versioned, schedulable, observable, and traceable capabilities (Zhang et al., 23 Dec 2025). NAIS similarly treats durable state, brokered access, governed execution, workflow orchestration, evidence generation, and scientist-in-the-loop review as core properties of an institutional research system (Huang et al., 13 Jul 2026).

3. End-to-end research workflows and representative systems

The general workflow begins with a human-supplied or system-derived research objective, followed by literature discovery and synthesis. A scientifically adequate retrieval system must identify relevant and irrelevant papers, methods, contradictory findings, evidence quality, terminology variants, research gaps, and claim-to-source relationships. Current systems often rely on keyword matching or shallow semantic retrieval and may fabricate citations or misclassify established concepts as novel.

Hypothesis generation should produce candidates that are novel, scientifically plausible, testable, relevant to the objective, and specific enough to support experimental design. Systems named in the literature include ResearchAgent, MOOSE-Chem, TAIS, ProtAgents, SciAgents, and The AI Scientist. The survey does not establish that these systems routinely generate experimentally validated, genuinely novel discoveries (Gridach et al., 12 Mar 2025).

Experiment planning translates a hypothesis into protocols, models, reagents or materials, parameters, controls, measurements, safety constraints, expected outcomes, and contingency plans. Examples include BioPlanner, CALMS, Coscientist, LLM-RDF, CRISPR-GPT, and MatPilot. Execution may involve code, simulators, molecular-design packages, robotic systems, laboratory instruments, or high-performance computing.

Representative domain systems include:

• Chemistry: Coscientist combines GPT-4 with web search, documentation analysis, code execution, and robotic automation; ChemCrow augments GPT-4 with 18 chemistry tools; Organa performs robotic chemistry operations; LLM-RDF uses six specialized agents for reaction development; and MOFGen combines language generation, diffusion-based crystal generation, quantum-mechanical validation, and synthetic-feasibility analysis.
• Biology and biomedicine: BIA generates bioinformatics workflows; CellAgent uses Planner, Executor, and Evaluator agents for single-cell RNA-seq analysis; TAIS supports genetic-research planning but requires human validation; ProtAgents combines LLMs, reinforcement learning, structure prediction, docking, and physics; and KOSMOS generates radiation-biology hypotheses that require independent statistical auditing (Nusrat et al., 17 Nov 2025).
• Materials science: LLaMP combines retrieval-augmented generation with Materials Project data and atomistic simulation; MatPilot supports human–machine materials research; AtomAgents uses retrieval, simulation, and analysis for alloy design; and TopoMAS combines dynamic knowledge graphs with first-principles calculations.
• Physics and astronomy: StarWhisper supports observation planning, telescope control, transient detection, and follow-up proposals; Mephisto uses spectral-energy-distribution fitting and tree search for JWST interpretation; and OpenFOAMGPT converts natural-language descriptions into computational-fluid-dynamics workflows.
• General machine-learning research: AI Scientist-v2 generates ideas, constructs baseline code, performs hyperparameter tuning, executes research modifications, conducts ablations and replications, and drafts manuscripts through progressive agentic tree search (Yamada et al., 10 Apr 2025).
• Pharmaceutical research: DiscoVerse decomposes archival drug-development questions across preclinical, clinical, and strategic branches, uses hybrid retrieval and cross-document linking, and produces source-linked reverse-translation analyses over confidential pharmaceutical archives (Zheng et al., 23 Nov 2025).
• Biomedical institutional workflows: NAIS orchestrates protected-data cohort extraction, phenotype construction, GWAS execution, quality control, reruns, and publication-oriented outputs while keeping raw genotype and EHR data inside institutional boundaries (Huang et al., 13 Jul 2026).

The systems differ in the location of autonomy. A laboratory robot may execute experiments autonomously but operate toward a human-specified target. A literature agent may search broadly but lack causal interpretation. A computational scientist may generate and test hypotheses but remain tied to a fixed dataset and metric. Agency is therefore multidimensional rather than binary.

4. Empirical demonstrations and evaluation

Evaluation of Agentic Scientists must extend beyond language quality and single-turn task accuracy. Relevant dimensions include scientific novelty, correctness, reproducibility, experimental success, autonomy, efficiency, calibration, reliability, safety, human usefulness, causal quality, and real-world scientific impact.

AI Scientist systems

An independent evaluation of Sakana’s AI Scientist found that all ten generated ideas and two seed ideas were classified as novel, including established concepts such as micro-batching for stochastic gradient descent, adaptive learning rates, hybrid matrix factorization, and e-fold cross-validation. Of twelve attempted experiments, five failed because of unresolved coding errors, corresponding to approximately 42%. The successful experiments produced seven manuscripts, but four contained incorrect or hallucinated numerical results and four contained structural defects. Manuscripts had a median of five citations, with only five of 34 references from 2020 or later. The reviewer agent rejected all seven AI-generated manuscripts but failed to identify several serious flaws, including invalid baselines, hallucinated numbers, repeated sections, and flawed experimental logic (Beel et al., 20 Feb 2025).

AI Scientist-v2 removed reliance on human-authored code templates, introduced progressive agentic tree search, added a dedicated experiment manager, and integrated vision-language-model review of figures and manuscripts. Three fully autonomous manuscripts were submitted to an ICLR workshop; one received reviewer scores of 6, 6, and 7, averaging 6.33 out of 10 and exceeding the workshop’s human acceptance threshold. The accepted paper nevertheless contained an inaccurate figure caption, a mismatch between a figure and its textual interpretation, and approximately 57% overlap between training and test examples in one data-generation setup. The paper was withdrawn after review (Yamada et al., 10 Apr 2025).

Jr. AI Scientist begins with a human-selected baseline paper, its LaTeX source and PDF, and its implementation repository. It analyzes limitations, proposes improvements, modifies multi-file code, runs experiments, performs ablations, and writes a paper. Its DeepReviewer average rating was 5.75, compared with 4.50 for Zochi, 3.92 for CycleResearcher-12B, 3.30 for AI Scientist-v1, 3.25 for AI Researcher, and 2.75 for AI Scientist-v2. However, only three selected papers were evaluated, without statistical significance analysis. Author-led inspection found ambiguous method descriptions, unsupported figure interpretations, irrelevant citations, and descriptions of experiments that had not occurred. In one case, invalid batch normalization created a misleading performance improvement (Miyai et al., 6 Nov 2025).

Closed-loop specialized systems

SR-Scientist elevates an LLM from equation proposer to an agent that analyzes data, writes executable equation code, submits equations for evaluation, receives numerical feedback, and revises its hypotheses. On synthetic symbolic-regression problems across chemistry, biology, physics, and materials science, it reported absolute gains of approximately 6% to 35% over relevant baselines. With GPT-OSS-120B, it reached accuracy-to-tolerance values of 63.57% at 1% tolerance and 49.35% at 0.1% tolerance. Removing data analysis and experience memory degraded performance, demonstrating the value of empirical inspection and long-horizon state (Xia et al., 13 Oct 2025).

KOSMOS generated three radiation-biology hypotheses. Independent reconstruction classified one as supported, one as plausible but uncertain, and one as false or unsupported. The supported association involved CDO1 expression and an induced radiation-response module, with $r=0.70$ and empirical $p=0.0039$. A 12-gene prostate-cancer signature achieved a concordance index of 0.613 with empirical c-index $p=0.0166$, but its effect size was not unique against random signatures. A DDR score predicting p53 response yielded $\rho_s=-0.40$ and empirical $p=0.756$. The study demonstrates that fluent, biologically plausible hypotheses require falsification against appropriate empirical null models (Nusrat et al., 17 Nov 2025).

AI CFD Scientist integrates literature-grounded ideation, OpenFOAM execution, source-code modification, mesh-independence checks, VLM-based physics verification, reference comparison, and figure-grounded manuscript writing. In an open-ended periodic-hill task at $Re_h=5600$, it discovered a Spalart–Allmaras runtime correction that reduced lower-wall $C_f$ RMSE from 0.004297 to 0.003958, a reported reduction of 7.89%. A planted-failure study found that the VLM gate detected 14 of 16 silent failures, or 87.5%, but produced four false positives, yielding a precision of 77.8% (Somasekharan et al., 7 May 2026).

AICS divides control design into task modeling, controller synthesis, and numerical tuning. On ControlEval, it achieved pass rates of 98% for first- and second-order stable systems, 100% for first-order unstable systems, 90% for second-order unstable systems, 100% for delay systems, and 74% for higher-order systems. Its reported average iterations ranged from 1.00 to 2.44 for these categories. Removing numerical optimization substantially reduced performance, supporting the separation between LLM-based search-space construction and external numerical optimization (Wang et al., 27 Aug 2026).

Domain knowledge and modality selection

A controlled property-insurance experiment showed that generic tabular modeling can fail even when the decisive information is present in associated data. A tabular-only random forest achieved normalized Gini 0.3823, whereas adding full CLIP image features achieved 0.7719, using a vision-language-extracted roof-health label achieved 0.7271, and supplying the true latent label achieved 0.8310. The oracle, which also knew the exact loss-generating mechanism, achieved 0.8379. The experiment indicates that agentic data science must discover relevant variables and modalities, not merely optimize models over an existing table (Luo et al., 24 Dec 2025).

5. Reliability, validity, and governance

The principal failure modes of Agentic Scientists are epistemic rather than merely syntactic.

Novelty failure occurs when a system mistakes established concepts for new ideas. Keyword-based retrieval may miss terminology variants, historical origins, or methodological equivalence.

Experimental-validity failure occurs when code executes but does not test the intended hypothesis. Examples include fixed parameters that should vary, omitted baselines, invalid comparison conditions, data leakage, inappropriate normalization, and resource metrics that do not support the claimed conclusion.

Interpretive failure occurs when the system reports a numerical result but assigns it an unsupported meaning. A one-second runtime reduction does not establish energy efficiency if memory use increases and no uncertainty or energy-accounting protocol is provided.

Citation and provenance failure occurs when papers are outdated, irrelevant, incorrectly attributed, or disconnected from the claims they are used to support.

Self-correction failure occurs when agents repair runtime errors without detecting conceptual errors, repeat a flawed assumption, optimize a proxy rather than scientific validity, or interpret a failure as noise rather than evidence against a hypothesis.

Calibration failure occurs when confidence is not aligned with correctness. This is particularly consequential in biomedical, chemical, biological, and physical-laboratory settings.

Error propagation occurs when unsupported retrieval becomes flawed synthesis, which becomes an invalid hypothesis, inappropriate experiment, mistaken analysis, and false report.

Governance systems address these problems through source-linked evidence, immutable or versioned artifacts, explicit experiment logs, sandboxed execution, access controls, parameter bounds, human approval, replication, independent checkers, and claim-to-evidence mappings. DiscoVerse demonstrates provenance-preserving archival synthesis; NAIS demonstrates broker-mediated access to protected biomedical data; and Bohrium plus SciMaster emphasize versioned capabilities, execution traces, resource accounting, validation gates, and reusable workflow artifacts.

The infrastructure problem includes resource discovery. A hierarchical MCP-server architecture allows an agent to negotiate across cloud, edge, and HPC environments through worker-level Secretary Agents and provider manifests. Across 19,973 negotiation simulations, the system reported 87.71% accuracy, with the largest errors involving missed software versions, storage, network requirements, and container mismatches. Agentic selection required approximately two seconds, compared with fractions of a millisecond for deterministic heuristics, although the agent could reason over more complex resource preferences (Sochat et al., 5 Aug 2026).

Privacy and institutional governance are essential when systems operate over clinical, genomic, pharmaceutical, or confidential laboratory data. NAIS restricted the agent to approved HTTP actions, prevented direct access to raw EHR and genotype data, returned only aggregate artifacts, and required human approval of SQL plans and phenotype definitions. Its hypertension GWAS involved 286,422 individuals and reproduced established loci including FGF5, ATP2B1, CNNM2, FTO, and GRB14, but the phenotype discrepancy required human clinical judgment. The system exposed and operationalized the discrepancy; it did not independently determine the scientifically appropriate phenotype (Huang et al., 13 Jul 2026).

A further controversy concerns whether interpretation itself can be delegated. The position that agentic research is “oxymoronic” argues that computation may be delegated when independently checkable, whereas interpretation requires human understanding and responsibility. This view is normative rather than an empirical finding, but it identifies a genuine distinction between reproducible execution and reproducible scientific judgment (Hogg, 31 Aug 2026).

6. Future directions and the status of the concept

A robust Agentic Scientist would require capabilities not yet demonstrated together across diverse domains:

1. Value-sensitive problem selection: selecting problems according to scientific significance, practical bottlenecks, societal need, feasibility, safety, and expected information gain rather than merely optimizing measurable proxies.
2. Tacit and failure knowledge: representing procedural adjustments, negative results, abandoned hypotheses, anomalous observations, instrument artifacts, and reasons for distrusting apparently successful experiments.
3. Persistent epistemic state: maintaining explicit hypotheses, predictions, uncertainties, failed attempts, measurement models, objective changes, provenance, and belief updates across long investigations.
4. Diversity-preserving hypothesis generation: avoiding convergence toward consensus caused by training distributions and preference optimization. Multiple models should not be assumed to provide independent epistemic samples when their hypotheses are highly overlapping.
5. Falsification-oriented experimentation: selecting experiments that discriminate among competing hypotheses, include controls and replication, and return rich evidence rather than only scalar rewards.
6. Multimodal and causal reasoning: integrating text, tables, figures, molecular structures, sequences, spectra, microscopy, sensor data, simulations, and laboratory metadata while distinguishing correlation, mechanism, and causation.
7. Simulation and physical verification: using scientific simulators as training and verification environments while accounting for model discrepancy and sim-to-real failure. Physical experiments remain necessary for claims about the world.
8. Reproducible execution and attribution: recording prompts, model versions, code, parameters, tool calls, artifacts, failures, environmental conditions, and human interventions. Proposed systems include Research Process Markup Language and Research Attribution Markup Language for workflow and contribution tracking.
9. Independent auditing: employing replication agents, code-result consistency checks, manuscript-to-log verification, appropriate null models, statistical safeguards, and human domain review.
10. Governed infrastructure: exposing scientific tools through stable interfaces with explicit inputs, outputs, costs, failure modes, safety envelopes, access controls, and replayable traces.
11. Human accountability: retaining human authority over objectives, ethical constraints, ambiguous results, high-risk actions, scientific significance, interpretation, and publication.

The current evidence supports a qualified characterization. Agentic systems can orchestrate research-shaped workflows, generate executable artifacts, adapt to feedback, discover candidate improvements, and reduce the time and cost of routine scientific work. They can function as research assistants, research-engineering systems, domain-specific co-scientists, or governed workflow orchestrators. They do not yet reliably demonstrate open-ended autonomous scientific intelligence.

The Agentic Scientist is therefore best understood as a spectrum of systems combining procedural autonomy with varying degrees of epistemic autonomy. The former—sequencing operations, invoking tools, modifying code, executing simulations, retrieving evidence, and drafting reports—is already demonstrated in constrained settings. The latter—choosing what matters, generating genuinely diverse and non-obvious hypotheses, learning from anomalous failure, revising objectives, establishing causal explanations, and producing independently validated knowledge—remains incomplete.

The central unresolved issue is not whether an AI system can produce something that resembles a scientific paper. It is whether the system can participate reliably in the epistemic norms of science: novelty grounded in prior work, evidence-sensitive belief revision, falsification, reproducibility, uncertainty calibration, transparent provenance, and accountable interpretation. Until these properties are demonstrated across diverse domains and independently verified, the Agentic Scientist remains an important research paradigm and a powerful supervision-dependent collaborator rather than a replacement for scientific judgment.