Reinventing the Python Notebook with Akshay Agrawal

What It Covers

Akshay Agrawal, creator of Marimo and former Google Brain researcher, explains how his open-source reactive Python notebook solves core Jupyter limitations — hidden state, poor version control, and the research-to-production gap — by storing notebooks as pure Python files with a static dependency graph.

Key Questions Answered

• •Reactive execution model: Marimo builds a static dependency graph by analyzing variable definitions and references in each cell using Python's built-in AST module. When a cell runs, all downstream cells automatically re-execute or are marked stale. This eliminates hidden state bugs that plague Jupyter users, where deleted cells leave ghost variables in memory for hours.
• •Pure Python file format: Marimo stores notebooks as standard `.py` files where each cell is a decorated function. This enables direct Git versioning without base64-encoded binary blobs, running notebooks as CLI scripts with `python mynotebook.py`, and importing functions directly from notebooks into other Python modules using standard import syntax.
• •Reproducible environments via UV and PEP 723: Marimo integrates with the UV package manager to auto-detect imports, prompt installation, and record dependencies in a PEP 723-compliant comment block at the file's top. On next run, UV creates an isolated virtual environment with exactly those packages, making environment reproduction a zero-configuration operation.
• •Notebook-to-web-app pipeline: Any Marimo notebook converts to a read-only interactive web app using `marimo run mynotebook.py` from the CLI, hiding code cells for non-technical stakeholders. UI elements like sliders, dropdowns, and selectable scatter charts bind directly to Python variables through the same reactive graph, requiring no callbacks or separate frontend framework.
• •LLM-friendly architecture: Because Marimo notebooks are plain Python files, Claude and other LLMs generate them reliably and can run them as scripts to verify correctness. Marimo's built-in AI assistant also accepts tagged variable references — like a DataFrame — and automatically injects schema and sample values as context, producing data-specific code generation rather than generic outputs.