Agentic Data Science in a Reproducible Marimo Notebook

Dan and Eric Ma discuss what changes when a data scientist treats coding agents as collaborators in an exploratory workflow. Eric describes how quickly his own ideas became obsolete: an agentic data science skill built around Markdown journals and generated plots gave way within weeks to Marimo Pair, where an agent can work directly inside a live notebook. Eric uses Marimo notebooks for more than interactive analysis. They preserve debugging sessions beyond an agent's context window, give colleagues a shareable pre-read, and keep code, visualizations, and environment definitions together in a single Python file. The live demonstration revisits a protein-engineering analysis from Eric's time at Novartis. An agent pulls activity and stereoselectivity data from a published paper, filters single-point mutants, builds ECDFs and a joint distribution, and maps activity or selectivity across protein positions. As we see live, agents don't always succeed. The protein-structure viewer fails during the demo, and Eric explains that an agent can debug its way through the problem but does not always finish within the available context. Even so, the ability to attempt a custom 3D visualization changes what feels feasible. Eric says he would previously have avoided the task or spent hours wrangling a PDB file and old PyMOL documentation. From there, the conversation moves to guardrails. Eric uses an Arrow of Intent approach to carry a project's high-level design through feature-level documents and testable specifications, giving both an agent and a future version of himself a record of what the software is supposed to mean. He pairs that with opinionated repository templates and AGENTS.md instructions about refactoring, documentation, testing, and coding style. The final turn is about attention. Eric finds outcome-driven agentic coding less tiring than solving every technical problem by hand, but context switching between foreground tasks caused burnout. His current practice is one demanding foreground task, two background tasks, and as many unattended automations as possible. The point is not to control every step an agent takes. It is to define the output, preserve intent, and build enough structure that the work remains understandable when the context window, model, or human changes. Chapters: 00:00 Why Eric's data science workflow changed so quickly 05:49 Marimo Pair puts agents inside the notebook 07:19 Reproducing a protein-engineering paper 10:45 Why enzyme stereoselectivity matters 13:20 Canonical data sources and portable environments 18:08 Exploring activity and selectivity with interactive plots 27:47 Finding promising regions for protein engineering 30:24 From heat maps to a 3D protein viewer 35:54 Learning from agent traces and preserving intent 47:13 Opinionated structure for agent-written code 52:25 Managing attention across agentic tasks