By Camilo Pedraza Jiménez, Applied Data Scientist at Probabl
Katerina Michenina, Market Intelligence Analyst at Probabl
Cailean Osborne, Head of Ecosystem Development at Probabl
This is the second post in our series of data deep dives into agentic data science trends.
In our first data deep dive, we found that downloads of core Python data science libraries have grown significantly faster since AI coding agents arrived on the market. Scikit-learn alone was downloaded 2 billion times in the last year, a 93% increase from the year before. We hypothesized that this download boom is, in large part, driven by agents. We can’t prove it causally but we found positive, statistically significant, and strengthening correlations between the downloads of core data science libraries like scikit-learn and the SDKs of market-leading AI agents like Anthropic's Claude Code and OpenAI's Codex.
That leaves us with the question: if data scientists are using AI agents, which ones are they using and at which stages of their workflows? In this deep dive, we provide data-driven answers to this question. Our main findings are:
Let’s quickly recap what “agentic data science” means. Our CSO Gaël Varoquaux and our CPO Yann Debray define agentic data science as the practice of orchestrating and overseeing AI agents that execute loops of iterations through data science workflows and machine learning experiments [1]. This definition builds on Andrej Karpathy's framing of agentic engineering as the practice of orchestrating agents that do 99% of the coding and overseeing things like correctness, security, taste, and maintainability [2].
By this logic, a “Data Scientist Agent” is a system autonomous enough to scope, execute, and validate an end-to-end data science workflow with minimal human intervention [3]. Again, consider the analogy of how coding agents like Claude Code or Cursor are used by engineers. For example, Anthropic’s analysis of roughly 400,000 Claude Code sessions showed that users mostly make the planning decisions (what to do) while the agent executes the engineering. By that bar, a Data Science Agent would be one system that owns an end-to-end data science workflow the way Cursor or Claude Code owns a coding task [3].
In “The path toward Data Scientist Agents that enterprises can trust in production” [3], Gaël and Yann discussed the key challenges around delegating end-to-end data science workflows to agents, and identified the types of tasks and decisions that make agentic solutions more or less suitable at 7 discrete stages of the data science workflow (see Figure 1). In a nutshell, what it boils down to is the kinds of contextual knowledge, statistical scrutiny, and judgment calls required at certain stages. For example, training a model against a loss curve can be automated: the metric tells the agent when it's done. However, knowing whether AI-generated code follows proper methodology or deciding whether a missing value is noise is hard to automate: no metric captures that context.
Judgment and scrutiny are crucial in such stages. As Eric Ma put it on his blog: AI tools can build but can't decide what's worth measuring [4]. Similarly, Hugo Bowne-Anderson argued on his podcast Vanishing Gradients that without the necessary guardrails (e.g. skills, a statistical harness) we run the risk that agents produce flawed models and that plausible-but-wrong conclusions inform business decisions [5]. Given these challenges, to the best of our knowledge, there is no Data Scientist Agent yet that can orchestrate a complete data science workflow end-to-end and let us trust the outcomes.
Figure 1: Seven stages in the agentic data science workflow
In this data deep dive, we build on Yann and Gaël's posts by taking a look at which agentic solutions exist for typical tasks involved in each distinct stage of the data science workflow and how much they are being used.
Our hypothesis is that data scientists are adopting agentic solutions the most in the stages of their workflows where there's a clear optimization signal, for example model training (e.g. the loss curve) and feature engineering (e.g. downstream model performance). By the same logic, we expect lower adoption at stages where data scientists have to frequently make judgment calls or require contextual knowledge: data ingestion (e.g. trusting a source), data cleaning (e.g. imputing values), and model evaluation (e.g. choosing the right metric).
To test this, we mapped solutions into the seven workflow stages and zoomed in on the ones that have a dedicated Python library or SDK given that these come with publicly available download data that provide insights into adoption and popularity. Specifically, we investigated the download trends of these solutions since February 2025 (the launch of Claude Code preview), which we take as a rough start date of the agentic coding era. We additionally ran analyses for the last six months, from February 2026 until July 2026.
We identified 13 agentic solutions that data scientists are using in their workflows, and found that coverage of the data science workflow is near-complete. What's interesting is that each solution maps particularly well onto 1 or 2 stages of the workflow (e.g. feature engineering or model training), rather than offering general-purpose assistants for end-to-end workflow execution. For example, Marimo lands squarely in data pre-processing and feature engineering: its cells are built so that cleaning steps can't spoil the data downstream through one-too-many reruns, and its selectable dataframes turn a click on a chart into rows you can access programmatically. At the other end, a platform like Arize is built for what happens after deployment, tracking model behaviour in production to surface drift and degradation.
However, not every solution has a corresponding Python library or SDK. None of the solutions in the first two stages (data ingestion and data cleaning) have Python libraries. Model training and monitoring in production are the stages with the most identified solutions, three each. Note: skore lib is not an SDK but the open source library for Skore, Probabl's agentic platform for tracking, exploring, and sharing the results of machine learning experiments. Additionally, aideml was included in two stages of the workflow since it is an agentic solution developed for both model training and model evaluation.
Figure 2: Agentic solutions for each stage of the data science workflow
The data supports our hypothesis: Agentic solutions are being adopted most at the stages of the data science workflow where tasks have a clear optimization signal. In particular, the solutions for model training have the most downloads and the steepest growth from around 1,500 downloads in March 2025 to around 19 million in July 2026. See Figure 6 in the appendix for the downloads of all solutions.
As in deep dive #1, we used compound monthly growth rate (CMGR) to study download trends in more depth since the start of the agentic coding era. CMGR measures the average monthly percentage growth in downloads over a period. The stage-level CMGR is a weighted average [6], meaning that SDKs with higher download volumes and longer tracking histories count more.
Growth of downloads since the agentic era began is fastest in the stages where we hypothesized it would be. The downloads of solutions for model training are growing the fastest at 86.7%, followed by solutions for feature engineering at 18.0% and monitoring models in production at 12%. When we zoom into the last 6 months, we observe that solutions for model training are growing the fastest (34.9%), followed by solutions for monitoring in production (18%) and model evaluation (14.3%).
With regards to the individual Python libraries and SDKs for the full period, google-adk has the fastest cumulative monthly growth (95.8%), followed by marimo (18.0%), arize-phonix (13.7%), and skore lib (10.8%). Again when we zoom into the last 6 months, google-adk continued having the highest growth rate (35.4%), followed by arize-phoenix (18.2%) and skore lib (14.7%. See Figure 7 in the appendix for the rest.
Figure 4: Growth rates of Python libraries and SDKs for agentic data science solutions
We examined the correlations between scikit-learn downloads and the Python libraries for these solutions at the three stages where scikit-learn is used: feature engineering, model training, and model evaluation. We used the same detrended Spearman method as deep dive #1, which removes shared-trend effects [7]. To get enough data points for reliable significance testing, we used daily rather than monthly downloads for this analysis.
When looking at the data since the start of the agentic era, all five solutions show a positive, statistically significant correlation with scikit-learn: h2o (r = 0.74), google-adk (r = 0.71), marimo (r = 0.71), skore lib (r = 0.37), and aideml (r = 0.12). ml-intern was excluded because its PyPI download history only starts in late April 2026. We also analyzed the data for the last six months only, and the results still hold for all libraries, except for aideml, which still has a positive but non-significant correlation with scikit-learn.
How are data scientists actually using agentic solutions? We know data scientists are using agentic solutions more and more (for example, to train and evaluate machine learning models), but we still don’t know how precisely. The State of Agentic Data Science 2026 survey by Vanishing Gradients and PyMC Labs just launched to tackle this data gap. If you’re a data scientist who uses agents, please share your insights by completing the survey (takes 1 minute)!
The findings are statistically significant, but it's important to acknowledge the limitations of this analysis. Here’s our round-up:
[1] Debray, Y. & Varoquaux, G. (2026). We’re teaching agents the skills of data science: Methodology and statistical thinking. https://blog.probabl.ai/teaching-agents-data-science-skills
[2] Karpathy, A. [@karpathy]. (2025).There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. https://x.com/karpathy/status/2019137879310836075
[3] Debray, Y. & Varoquaux, G. (2026). The path toward Data Scientist Agents that enterprises can trust in production. https://blog.probabl.ai/data-science-agents-you-can-trust
[4] Ma, E. (2026). What data science is actually about in the age of AI. https://ericmjl.github.io/blog/2026/5/20/what-data-science-is-actually-about-in-the-age-of-ai/.
[5] Bowne-Anderson, H. & Wietecki, T. (2026). The Future of Agentic Data Science. Interview by H. Bowne-Anderson. Vanishing Gradients. https://hugobowne.substack.com/p/the-future-of-agentic-data-science.
[6] Chugani, V. (2025). Weighted Average Formula: A Complete Guide with Practical Applications. https://www.datacamp.com/tutorial/weighted-average-formula.
[7] Seifer, J. (2020). Averaging Correlations - Part II. https://medium.com/@jan.seifert/averaging-correlations-part-ii-9143b546860b.
Figure 6: Downloads of Python libraries and SDKs for agentic data science solutions
Figure 7: Growth rates of Python libraries and SDKs for agentic data science solutions