Coding agents reach into the running notebook kernel to drive analysis live

When a coding agent can read and write a running notebook kernel directly, data analysis stops being a prompt-and-paste loop and becomes a live pair-programming dialogue on a reactive canvas. The agent reaches into the kernel of a running Python notebook and manipulates it: it sees the variables in context, writes the next cell, and when a call errors it fixes itself in a loop because it can observe the kernel state directly.

Marimo is the canvas that makes this work because it is fully reactive. Change one cell and every downstream cell recomputes, so the agent never trusts a stale output the way a Jupyter user can. Its sandbox mode carries dependencies per-notebook through PEP 723 inline metadata, so the agent can auto-install what each cell needs and the resulting notebook is portable. The human sets direction step by step (see: break work into small iterative chunks) while the agent implements, and the tight kernel feedback sets the pace (see: the rate of feedback is an agents speed limit).

A second-order effect: the tedious parts of notebooks get cheap. No one writes markdown documentation cells by hand because they are tedious, but an agent that already holds the kernel state can dictate them, interpolating the notebook’s own variables into a readable summary. The output is a notebook a colleague can actually read, not just code a data scientist ran.