The Super Data Scientist: An Agentic Architecture for Expanding Human Intelligence
Why the value of AI agents is not only analytical speed, but a wider proposition space, deeper inquiry, and stronger decision intelligence.
Most discussions about AI in data science begin with productivity.
AI can write SQL, generate Python code, clean data, select models, create visualizations, and draft reports. These capabilities are valuable, but productivity is only the first-order effect.
The deeper transformation is epistemic.
AI agents can expand the range of propositions a data scientist can formulate, the bodies of knowledge that can be consulted, the methods that can be considered, the evidence that can be examined, and the depth with which competing explanations can be tested.
A data scientist equipped with a properly designed system of AI agents does not simply perform the same work faster. That scientist can operate beyond the normal boundaries of individual memory, disciplinary training, technical specialization, and available time.
This creates the possibility of a new professional form:
The super data scientist is a human intelligence architect who orchestrates AI agents, analytical methods, domain knowledge, and scientific controls to explore a broader proposition space, conduct deeper investigations, and connect validated evidence to consequential decisions.
This is not a claim that AI makes one person omniscient. Nor is it a claim that agents can replace statistical expertise, domain knowledge, or scientific judgment.
It is a claim about system-level intelligence.
The relevant unit of analysis is no longer the isolated data scientist or the individual AI model. It is the combined system:
Intelligence system = human judgment × agentic exploration × analytical execution × adversarial validation × domain context.
If any component is weak, the entire system becomes unreliable. If the components are carefully designed and integrated, the system can produce a form of analytical capacity that neither the human nor the AI could achieve independently.
1. The traditional boundary of a data scientist
Every data scientist operates within an invisible intellectual boundary.
That boundary is shaped by:
- formal education
- professional experience
- preferred analytical methods
- familiarity with particular industries
- knowledge of available data
- exposure to relevant theories
- access to collaborators
- time available for investigation
- cognitive capacity for comparing alternatives
A data scientist trained primarily in predictive modeling may naturally frame a business problem as a classification or regression problem. A statistician may emphasize estimation and uncertainty. An econometrician may focus on causal identification. An operations researcher may formulate an optimization problem. A behavioral scientist may ask how incentives, perception, or bias shape the observed behavior.
Each perspective may reveal part of the truth.
The limitation is that an individual practitioner may not recognize all the relevant perspectives. We tend to ask questions that we already know how to answer. We tend to select methods that exist within our technical comfort zones. We search for evidence using the terminology of disciplines we already understand.
Consequently, many analytical projects are constrained before the first line of code is written.
The initial framing determines:
- which data is considered relevant
- which mechanisms are treated as plausible
- which hypotheses are tested
- which methods are selected
- which outcomes are optimized
- which risks remain invisible
AI agents can help weaken this constraint by introducing concepts, methods, evidence, and alternative explanations from outside the scientist’s original intellectual territory.
This is the first foundation of the super data scientist:
AI agents expand not only the speed of analysis, but the reachable space of inquiry.
2. From computational acceleration to cognitive expansion
AI assistance in data science can be understood as four increasingly powerful levels.
Level 1: Mechanical acceleration
The AI assists with code generation, syntax correction, data transformation, documentation, visualization, and report preparation.
The same analytical plan is executed more efficiently.
Level 2: Analytical acceleration
The AI proposes features, recommends models, performs diagnostic tests, evaluates alternatives, and automates experiments.
The analytical plan becomes larger and faster, but it generally remains within the scientist’s original conceptual framework.
Level 3: Cognitive expansion
The AI introduces unfamiliar theories, methods, analogies, causal mechanisms, research findings, and disciplinary perspectives.
The scientist can now consider questions and explanations that might not have emerged from unaided reasoning.
Level 4: Agentic intelligence
Multiple specialized agents participate in a structured cycle of proposition generation, evidence collection, methodological design, experimentation, criticism, synthesis, and decision evaluation.
At this level, AI is no longer merely a tool used within a workflow. It becomes part of the architecture through which the scientist thinks.
The current public conversation remains concentrated at Levels 1 and 2. The concept of the super data scientist begins at Levels 3 and 4.
3. The proposition space is the true frontier
A data science project is often described as a process of transforming data into insights.
That description is incomplete.
Before data can produce intelligence, someone must decide what propositions deserve investigation.
A proposition is a structured and testable claim about the world. A useful proposition should specify:
- the phenomenon being explained
- the proposed mechanism
- the relevant population
- the conditions under which the mechanism operates
- the observable evidence expected if the proposition is true
- the evidence that would contradict it
- the decision that could change if it is validated
Consider customer attrition.
A predictive question might be:
Which customers are likely to leave?
A proposition-oriented investigation asks richer questions:
- Customers leave because pricing complexity reduces trust.
- Customers leave because unresolved service interactions accumulate.
- Customers leave because the onboarding process fails to establish product value.
- Customers leave because competitors target them during specific lifecycle transitions.
- Customers appear to leave because the organization’s definition of an active customer is incorrect.
- Customers with high predicted risk are not necessarily customers whose behavior can be changed.
Each proposition implies different data, methods, experiments, interventions, and business decisions.
Traditional data science often investigates a small number of propositions because proposition development is expensive. Each additional proposition requires literature review, data discovery, operationalization, analysis, validation, and explanation.
AI agents can reduce the marginal cost of proposition exploration.
They can help generate a portfolio of competing propositions, identify their assumptions, locate relevant evidence, design tests, and determine what observations would distinguish one explanation from another.
Research on AI-supported scientific hypothesis generation increasingly describes discovery as an iterative system involving knowledge integration, text and concept mining, simulation, causal inference, multi-agent exploration, and human validation. This work also identifies novelty, feasibility, transparency, bias, and cross-disciplinary reasoning as essential challenges. Scientific Hypothesis Generation and Validation (arXiv preprint)
Researchers at MIT developed SciAgents, a multi-agent framework that uses knowledge graphs and specialized agents to make connections across scientific concepts and generate research hypotheses. Its architecture attempts to reproduce some of the intellectual diversity normally created by communities of scientists from different backgrounds. MIT CEE: “Need a research hypothesis? Ask AI.”
Co-Scientist, developed by Google DeepMind, similarly uses specialized agents to generate, compare, criticize, and refine scientific hypotheses. The architecture treats discovery as a repeated cycle of generation, debate, reflection, and evolution, rather than as a single prompt followed by a single answer. Gottweis et al., “Accelerating scientific discovery with Co-Scientist” (Nature, 2026), Google DeepMind’s Co-Scientist overview
These developments suggest a more powerful interpretation of AI in data science:
The primary value of an agent is not that it produces an answer. Its primary value is that it enlarges and improves the population of propositions from which useful answers may emerge.
4. Breadth: moving beyond the original comfort zone
The first defining capability of the super data scientist is breadth of thought.
Suppose an organization observes declining employee productivity.
A conventional analysis might examine:
- working hours
- employee tenure
- compensation
- absenteeism
- manager ratings
- workload
- training participation
An agentic system could expand the investigation into additional perspectives:
- Measurement theory: Has productivity actually declined, or has the way it is measured changed?
- Organizational network analysis: Has collaboration become fragmented?
- Behavioral economics: Are incentives causing employees to optimize visible activity rather than real outcomes?
- Human factors: Did a new information system increase cognitive load?
- Operations research: Are bottlenecks outside the employee’s control reducing throughput?
- Causal inference: Did an organizational intervention cause the observed change?
- Labor economics: Have external market conditions altered retention or motivation?
- Sociology: Have trust, identity, or group norms changed?
- Process mining: Did the structure and sequence of work become more complicated?
- System dynamics: Is the apparent decline a delayed consequence of an earlier policy?
The objective is not to apply every perspective indiscriminately. The objective is to prevent the initial framing from prematurely restricting the investigation.
Evidence indicates that generative AI can help knowledge workers perform tasks beyond their established skill boundaries. In a controlled experiment involving 480 consultants and 44 data scientists, nontechnical consultants used AI to perform coding, statistical interpretation, and predictive-modeling tasks that would ordinarily fall outside their professional expertise. Sack et al., “GenAI Doesn’t Just Increase Productivity. It Expands Capabilities” (BCG, 2024)
A randomized trial reported in Nature also found that generative AI enabled knowledge workers to perform data science tasks outside their existing skill sets. However, much of that capability did not remain after the AI was removed. Wiles et al., “Generative AI and the temporary upskilling of knowledge workers” (Nature Human Behaviour, 2026)
This produces an important theoretical distinction:
- Internal expertise is knowledge and capability possessed by the human.
- Borrowed capability is performance made possible through AI assistance.
- System capability is the combined performance of humans, agents, tools, knowledge sources, and controls.
- Collective expertise emerges when human practitioners, AI agents, and qualified domain experts work together.
The super data scientist does not pretend that borrowed capability is personal mastery.
Instead, the super data scientist knows how to mobilize system capability while recognizing when genuine domain expertise is required.
5. Depth: investigating mechanisms, not merely patterns
Breadth creates more possible explanations. Depth determines whether any explanation deserves belief.
Suppose an analysis finds that customers who contact support frequently are more likely to leave.
A pattern-oriented analysis might conclude:
Frequent support interaction is a strong predictor of attrition.
A deeper investigation asks:
- Are dissatisfied customers more likely to initiate support contacts?
- Does support contact increase attrition, or is it a symptom of an earlier problem?
- Are repeated contacts evidence that the first interaction failed to resolve the issue?
- Does the relationship differ by issue type or customer tenure?
- Are some types of support interaction protective?
- Does the timing of the interaction matter?
- Is the effect caused by service quality, waiting time, communication style, or the underlying product defect?
- Are particular teams, channels, or workflows responsible?
- Does a policy change create a natural experiment?
- Which intervention would distinguish among the competing mechanisms?
A collection of specialized agents can pursue these branches in parallel:
- a causal agent constructs alternative causal diagrams
- a statistical agent tests robustness across specifications
- a temporal agent examines ordering and delayed effects
- a text-mining agent analyzes support conversations
- a process agent reconstructs the operational workflow
- a segmentation agent studies heterogeneous effects
- an economic agent estimates the value of possible interventions
- an adversarial agent searches for reasons the entire conclusion may be wrong
This does not eliminate the need for human reasoning. It gives human reasoning a larger field in which to operate.
The data scientist moves from asking:
What pattern can I model?
to asking:
What mechanisms could produce this pattern, what evidence would distinguish them, and which actions remain defensible under uncertainty?
This is the second foundation of the super data scientist:
The objective is not to maximize the number of analyses. It is to increase the depth of explanation.
6. Multiple agents should create intellectual opposition
A poorly designed agent tells the data scientist what the data scientist wants to hear.
A well-designed agentic system creates opposition.
Confirmation bias can enter a data science project through many apparently reasonable choices:
- selecting a favorable time window
- excluding inconvenient observations
- choosing a convenient target variable
- emphasizing one evaluation metric
- ignoring subgroup failure
- trying many specifications without correction
- treating predictive importance as causal importance
- searching more aggressively for confirming evidence
- interpreting weak evidence using confident language
A serious agentic architecture should therefore separate proposition development from proposition criticism.
For every major proposition, the system should construct:
A supporting case
What evidence is consistent with the proposition? Which mechanisms and studies support it? What analytical results strengthen it?
An opposing case
What evidence contradicts the proposition? Which alternative mechanisms explain the same observations? Which assumptions are fragile?
A methodological audit
Are the measurements valid? Is the sample representative? Is there leakage, selection bias, confounding, multiple testing, nonstationarity, or inappropriate model specification?
A business challenge
Even if the proposition is statistically valid, would acting on it produce meaningful value? Is the intervention feasible, reversible, ethical, and operationally sustainable?
This is not a cosmetic multi-agent conversation in which several agents produce similar summaries.
It is a designed system of criticism.
The agents should be rewarded for discovering:
- hidden assumptions
- weak identification
- unstable estimates
- conflicting evidence
- omitted populations
- unmeasured mechanisms
- implementation barriers
- unintended consequences
- conditions under which the recommendation fails
The super data scientist is not someone surrounded by machine assistants that agree.
The super data scientist directs a community of machine researchers that are required to disagree productively.
7. Analytical diversity must be designed
Multiple agents do not automatically create diverse reasoning.
If the agents rely on the same model, similar prompts, identical data, and the same evaluation criteria, their apparent debate may be a synthetic consensus.
Research published in Science Advancesillustrates this risk. Generative AI improved the evaluated creativity and quality of individual stories, especially for less creative writers, while making the resulting stories more similar to one another. Individual performance increased, but collective diversity decreased. Doshi and Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content” (Science Advances, 2024)
An equivalent risk exists in data science.
Five agents may generate five different reports while sharing:
- the same implicit assumptions
- the same popular theories
- the same data limitations
- the same dominant statistical paradigm
- the same source material
- the same blind spots
Agent diversity should therefore be architectural, not theatrical.
A strong system might introduce diversity through:
- different model families
- independent contexts
- distinct theoretical frameworks
- separate evidence collections
- statistical and qualitative methods
- predictive and causal perspectives
- internal and external data
- confirmatory and adversarial objectives
- human reviewers from different disciplines
- independent replication of consequential results
The purpose of a multi-agent architecture is not to create a louder consensus.
It is to create a more reliable process for discovering disagreement.
8. Intelligence requires movement across levels of abstraction
Many data science projects fail because they answer a technical question at the wrong level of abstraction.
A business asks:
Why are sales declining?
The analytical team produces a forecast.
The business asks:
Which customers should receive an intervention?
The team produces a risk score.
The business asks:
What action should we take?
The team presents feature importance.
These outputs may be mathematically correct, but they do not complete the path from evidence to decision.
A super data scientist must move across several levels of inquiry:
| Level | Primary question |
|---|---|
| Descriptive | What happened? |
| Diagnostic | Which conditions are associated with it? |
| Mechanistic | What process may have produced it? |
| Predictive | What is likely to happen next? |
| Causal | What would happen under an intervention? |
| Prescriptive | Which action should be selected under constraints? |
| Economic | Does the expected benefit justify the cost and risk? |
| Operational | Can the organization implement the action? |
| Ethical | Who benefits, who bears risk, and what should not be optimized? |
| Learning | How will the result update future decisions? |
AI agents can specialize at different levels while preserving the connections among them.
A predictive agent may identify customers at high risk of leaving. A causal agent should warn that high risk does not imply high persuadability. An optimization agent can allocate a limited intervention budget. A governance agent can identify inappropriate variables or customer groups. A monitoring agent can determine whether the intervention creates incremental value.
The intelligence of the system lies not in any one output, but in the integrity of the chain:
Observation → Explanation → Prediction → Intervention → Outcome → Learning
9. A reference architecture for the super data scientist
The wrong architecture is one general-purpose agent with unrestricted access to data, tools, and decisions.
The preferred architecture is a governed system of specialized agents with explicit responsibilities, evidence standards, and human approval gates.
Layer 1: Human intelligence and governance
The human data scientist acts as the intelligence architect.
Responsibilities include:
- defining the consequential decision
- setting the scope of inquiry
- establishing the standard of evidence
- specifying ethical and operational boundaries
- assigning analytical responsibilities
- resolving conflicts among agents
- involving domain experts
- approving conclusions and actions
- accepting accountability for outcomes
The human is not merely “in the loop.” The human provides intellectual leadership for the loop.
Layer 2: Knowledge and proposition formation
Horizon-scanning agent
Searches across disciplines, external developments, prior internal work, and structurally similar problems.
Proposition agent
Creates a portfolio of testable propositions containing mechanisms, conditions, observable implications, contradictions, and decision relevance.
Knowledge-graph agent
Maps relationships among concepts, datasets, studies, assumptions, methods, and organizational decisions.
Layer 3: Evidence acquisition and control
Evidence agent
Collects internal and external evidence while preserving source, date, scope, and provenance.
Data-quality agent
Examines lineage, missingness, duplication, measurement validity, temporal consistency, selection effects, and representativeness.
Context agent
Documents institutional, operational, legal, economic, and domain-specific conditions that affect interpretation.
Layer 4: Analytical execution
Statistical agent
Performs estimation, uncertainty assessment, hypothesis testing, sensitivity analysis, and model diagnostics.
Machine-learning agent
Develops predictive models, studies generalization, examines model stability, and evaluates performance across populations.
Causal agent
Constructs causal assumptions, identifies possible strategies, tests intervention effects, and separates association from causation.
Simulation and optimization agent
Evaluates scenarios, constraints, resource allocation, second-order effects, and possible intervention policies.
Layer 5: Criticism and validation
Adversarial agent
Attempts to invalidate the proposition and identify alternative explanations.
Reproducibility agent
Reconstructs the analysis from data, code, assumptions, and environment specifications.
Robustness agent
Tests sensitivity across samples, time periods, variable definitions, specifications, and model families.
Domain-review agent
Determines whether the conclusion is compatible with established domain knowledge and identifies issues requiring a qualified expert.
Layer 6: Decision and learning
Synthesis agent
Creates an evidence map that separates:
- strongly supported conclusions
- weakly supported conclusions
- disputed findings
- critical assumptions
- unresolved uncertainties
- high-value next investigations
Business-value agent
Estimates expected benefit, cost, downside risk, reversibility, implementation difficulty, and time to impact.
Decision agent
Constructs alternative actions rather than presenting one recommendation as inevitable.
Monitoring agent
Measures realized outcomes, unintended consequences, distribution shift, and deviations from original assumptions.
Learning agent
Updates the organization’s accumulated knowledge after interventions are implemented.
This architecture transforms a data science project into a continuing intelligence system.
10. The project lifecycle should begin with a decision, not a dataset
The conventional project lifecycle is often represented as:
Business question → data → model → insight → presentation
A super data scientist requires a more rigorous lifecycle:
Decision → proposition portfolio → evidence architecture → competing explanations → analytical tests → adversarial review → human judgment → intervention → outcome measurement → organizational learning
Phase 1: Identify the decision
Define:
- who must make the decision
- what action alternatives exist
- what outcomes matter
- what constraints apply
- what uncertainty currently prevents action
Phase 2: Construct the proposition portfolio
Generate explanations that differ meaningfully in mechanism, not merely in wording.
Each proposition should specify what evidence would support or weaken it.
Phase 3: Design the evidence architecture
Determine:
- what data already exists
- what data is missing
- how each variable was generated
- which external sources are relevant
- which evidence is independent
- which evidence is susceptible to bias
Phase 4: Conduct parallel but controlled exploration
Assign agents to investigate distinct propositions and methods. Preserve complete records of assumptions, code, sources, and analytical choices.
Phase 5: Apply adversarial validation
Require independent agents or human reviewers to challenge the data, methods, interpretation, and business relevance.
Phase 6: Integrate evidence
Do not simply average agent conclusions. Identify why conclusions differ and which assumptions explain the disagreement.
Phase 7: Make the decision
Present alternatives, expected consequences, uncertainty, reversibility, and monitoring requirements.
Phase 8: Learn from intervention
Compare expected and realized outcomes. Determine which propositions survived contact with reality.
Without the final phase, an organization generates reports but does not accumulate intelligence.
11. What must remain human
The emergence of agentic data science does not reduce the importance of expertise.
The AgentDS Technical Report evaluated AI-only systems and human-AI collaboration across 17 domain-specific challenges in commerce, healthcare, insurance, manufacturing, retail banking, and food production. AI-only baselines performed near or below the median of participating teams, while the strongest results came from human-AI collaboration. The report particularly identified domain-specific reasoning as a weakness of current agents. AgentDS Technical Report (arXiv, 2026)
Research on expert cognition similarly finds that experts are generally more willing to delegate repetitive information-foraging tasks than complex synthesis and interpretation. The researchers identify a fundamental tension between reducing cognitive effort and preserving the deliberate practice through which expertise develops. “Augmenting Expert Cognition in the Age of Generative AI” (arXiv preprint)
Therefore, several responsibilities should remain clearly human:
- deciding which problems deserve attention
- determining what outcomes are socially and organizationally valuable
- judging whether evidence is sufficient for action
- recognizing context that is absent from the data
- involving accountable domain experts
- resolving ethical conflicts
- accepting responsibility for consequential recommendations
The human contribution is not merely final approval.
It is the exercise of judgment throughout the architecture.
12. The greatest risk is verification debt
Agentic systems can generate propositions, analyses, code, and narratives faster than humans can validate them.
This creates verification debt.
Verification debt accumulates when the volume of generated analytical work exceeds the organization’s capacity to confirm:
- data validity
- methodological appropriateness
- reproducibility
- source reliability
- robustness
- domain consistency
- causal interpretation
- business relevance
The system may appear increasingly productive while becoming progressively less trustworthy.
Other important failure modes include:
Artificial breadth
The system mentions many disciplines but applies them superficially.
Synthetic consensus
Multiple agents agree because they inherit similar assumptions.
Hypothesis fishing
The low cost of analysis encourages testing until a desirable result appears.
Causal overreach
Sophisticated language disguises weak identification.
Context collapse
Findings are transferred across domains without respecting differences in population, institutions, measurement, or incentives.
Expertise atrophy
Excessive delegation reduces the scientist’s ability to recognize subtle errors.
Business theater
The system generates persuasive reports without changing measurable decisions or outcomes.
The solution is not less agentic capability.
The solution is stronger scientific architecture.
13. Becoming a super data scientist
The path does not begin with prompt engineering.
It begins with intellectual and methodological discipline.
Learn to formulate propositions
Convert vague business questions into claims with explicit mechanisms, assumptions, evidence requirements, and decision implications.
Strengthen statistical reasoning
Agents can execute a method without understanding whether the method addresses the real question. Knowledge of uncertainty, measurement, sampling, experimental design, causal inference, and validation remains foundational.
Develop cross-disciplinary literacy
The goal is not to master every discipline. It is to recognize when another discipline contains a useful conceptual model and when an expert must enter the system.
Learn systems architecture
Understand agents as components in an intelligence system:
- roles
- interfaces
- tools
- memory
- context
- orchestration
- evaluation
- observability
- permissions
- human approval gates
Practice adversarial thinking
For every conclusion, ask:
- What alternative mechanism explains the same result?
- What is the strongest evidence against this claim?
- Which assumption carries the greatest decision risk?
- What observation would cause us to change our conclusion?
- Which group or condition is missing from the analysis?
Preserve independent capability
Inspect raw observations. Reproduce consequential analyses. Read primary sources. Understand the assumptions behind important methods. Do not allow AI-generated fluency to substitute for comprehension.
Measure decisions, not analytical artifacts
The ultimate product of data science is not a model, notebook, dashboard, or report.
It is a better decision and a measurable outcome.
14. The enterprise advantage is an intelligence architecture
Access to capable AI models will become increasingly common.
The durable advantage will not come from possessing an agent. It will come from building a superior system of inquiry.
A mature organization should be able to answer:
- How are important propositions generated and recorded?
- Can evidence be traced to its source?
- Are conflicting explanations preserved?
- Are analytical assumptions machine-readable and human-reviewable?
- Can consequential findings be reproduced?
- Are agents evaluated by scientific and business criteria?
- Is disagreement rewarded?
- Are outcomes from implemented decisions returned to the knowledge system?
- Does each project improve the next project?
An organization with powerful AI but weak epistemic practices may simply become wrong more quickly and persuasively.
An organization with disciplined analytical leadership and a strong agentic architecture can create a compounding intelligence system.
Each project contributes:
- tested propositions
- reusable methods
- validated evidence
- rejected explanations
- decision outcomes
- newly discovered boundary conditions
Over time, the organization does not merely accumulate data.
It accumulates structured intelligence.
Conclusion: From data analysis to intelligence architecture
The golden age of data science will not be defined by faster production of code, models, dashboards, or reports.
It will be defined by the expansion of what a thoughtful data scientist can investigate.
AI agents can help a scientist move beyond familiar methods, enter neighboring disciplines, formulate a larger and more diverse proposition portfolio, conduct deeper investigations, institutionalize criticism, and connect evidence to business decisions.
But agents do not automatically create intelligence.
Intelligence emerges from the architecture that connects:
- human curiosity
- scientific reasoning
- agentic exploration
- methodological rigor
- domain expertise
- adversarial validation
- organizational action
- feedback from reality
The super data scientist is therefore not a data scientist who knows everything.
The super data scientist is someone who knows how to design, direct, criticize, and govern a system that can investigate far more than one person could investigate alone.
**A traditional data scientist analyzes the data placed in front of them.
A super data scientist expands the space of propositions, evidence, methods, and decisions that the organization is capable of considering.**
This is not merely faster data science.
It is a new architecture for human intelligence.
Further reading
- When an A/B Test Can Stop Early Without Lying
- An LLM Is Not a Control Group: The Statistical Problem with Synthetic A/B Tests
- Read the Chinese translation.
References
- Kulkarni, Adithya, et al. Scientific Hypothesis Generation and Validation: Methods, Datasets, and Future Directions. arXiv preprint, 2025.
- MIT CEE: Need a research hypothesis? Ask AI.
- Gottweis, Juraj, et al. Accelerating scientific discovery with Co-Scientist. Nature, 2026.
- Google DeepMind. Co-Scientist: A multi-agent AI partner to accelerate research.
- Sack, Daniel, et al. GenAI Doesn’t Just Increase Productivity. It Expands Capabilities. Boston Consulting Group, 2024.
- Wiles, Emma, et al. Generative AI and the temporary upskilling of knowledge workers. Nature Human Behaviour, 2026.
- Doshi, Anil R., and Oliver P. Hauser. Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 2024.
- Luo, An, et al. AgentDS Technical Report: Benchmarking the Future of Human-AI Collaboration in Domain-Specific Data Science. arXiv, 2026.
- Siu, Alexa, and Raymond Fok. Augmenting Expert Cognition in the Age of Generative AI: Insights from Document-Centric Knowledge Work. arXiv preprint, 2025.