marimo-pair
Markdown instructions plus a bash script that let a coding agent reach into the live Python kernel. The bridge Jupyter never had. Install and point it at your own work.
Show Us Your Agent Skills / EP 02 / guest dossier
ERIC MA
MODERNA
MARIMO PAIR
AGENTIC EDA
SOCRATES NOT SYCOPHANT
ERIC MA
Eric does data science as a live conversation: the agent edits a reactive Marimo notebook, reaches into the Python kernel, and renders plots, while he chooses every scientific question and owns the interpretation. Asking an agent to "do it all for me," he says, is irresponsible as a data scientist.
1 PLOTat a time: the analysis grows one artifact per question
0stale cells: reactive notebooks never go stale like jupyter
3Dprotein view: best mutations sat outside the active site
1/8the price: opus-quality models via a GLM coding plan
EVERYclaim needs an artifact that backs it
BY AGENTgit commits and pushes. he's stopped using the terminal
EP 02 · ERIC MA · agentic EDA in a live Marimo notebook, on stream
DATA SCIENCE AS CONVERSATION
the agentic-eda write-up. fork it
"I don't go into the analysis with a vague question and just ask the agent to do it all for me. That, I think, is irresponsible as a data scientist." So Eric splits the work: the agent handles the routine loops, and the human keeps "load the data context into my head" firmly in the loop. The agent is "a pair programmer, not a thing that just does the whole thing for me."
The enabler is the Marimo Pair skill: markdown instructions plus a bash script that let the coding agent reach directly into the Python kernel. The surface is a reactive Marimo notebook, where cells never go stale the way they do in Jupyter, so the agent can edit live while state stays coherent.
Inside the write-up: the principles below, a session walkthrough, anti-patterns, the finished notebook, and the AGENTS.md rules that keep it presentation-ready.
"It can help me amplify where I don't know what I don't know."
Eric's favorite thing about agents: not just doing the boring work, but surfacing the gaps. "There are ways to use AI as Socrates instead of a sycophant." 00:05:36
ONE PLOT AT A TIME
the live protein-engineering analysis. every timestamp opens the segment
Teach the domain firstBefore any plot, Eric explains the mutation notation: A.111.C means wild-type A at position 111, mutated to C. The agent can't plot what it can't read. 00:17:39
Ask for one plot, then correct itA heat map over single-point mutations, then a fix: Viridis for activity (0 to 1), divergent scale for chirality (-1 to +1). The scale changes the interpretation. 00:25:20
Leave notes, demand artifacts"You have to have artifacts that back the claims that you have." Eric writes a note to self about the weak chirality-activity relationship as the agent builds. 00:28:20
Build a custom view when plots run outAn AnyWidget + 3Dmol.js protein viewer: "coding agents are better at TypeScript than I am," so custom scientific interfaces get cheap. 00:32:09
Correct the agent's mistakes in plain language"Opus made a mistake on this one": it drew balls-and-sticks for everything and missed the ribbon. Eric fixes it by talking, without leaving the notebook. 00:32:51
Let the notebook write the conclusionThe agent calculates and interpolates the quantitative lack of correlation into a markdown cell. The finding: best mutations sat outside the active site. 00:37:54
"Load the data context into my head, versus this routine thing that I need a machine to automate."
Eric's split of data science into two buckets. EDA lives almost entirely in the first: looking, staring, understanding. The agent gets the second. 00:24:19
WHAT MAKES IT WORK
the pieces behind a notebook that stays readable and honest
agentic-eda
The full write-up: human directs one question at a time, every claim backed by an artifact, the agent as pair programmer.
AGENTS.md
Two rules keep the notebook presentation-ready: interleave explanatory markdown with code, and give every cell a unique descriptive name.
demo.py
The finished notebook from the live run: plots, widgets, the 3D viewer, notes to self, and the agent-written conclusion, all preserved.
"I do commits and pushes by agent commands. I don't do git commit on the terminal anymore."
His answer to what would embarrass him if his agent logs leaked. The boring work, including the ugly rebase edge cases, is the agent's now. 00:09:50
THE STACK
cursor + marimo + opus drive the demo. the rest is his hunt for cheaper intelligence