End-To-End Agent-Assisted Data Analysis Workflow Packaging

Issue 75 Edition 2026-03-16 5 min read

General

Sources: 1 • Confidence: High • Updated: 2026-03-17 15:15

Key takeaways

A three-hour NICAR 2026 workshop titled "Coding agents for data analysis" was delivered for data journalists, and a handout was prepared for it.
Total Codex token spend by workshop participants was 23 US dollars.
A highlighted workflow configured Datasette to serve static content from a visualization folder, then used Claude Code to iteratively create interactive visualizations directly in that folder.
The handout was designed to be useful to people who did not attend the workshop, and the author expects it to apply beyond data journalism to anyone exploring data.
The workshop demonstrated using Claude Code and OpenAI Codex to explore, analyze, and clean data.

A three-hour NICAR 2026 workshop titled "Coding agents for data analysis" was delivered for data journalists, and a handout was prepared for it.
The workshop demonstrated using Claude Code and OpenAI Codex to explore, analyze, and clean data.
The handout covered setup for Claude Code and Codex; asking questions against a database; exploring and cleaning data; creating visualizations; and scraping data with agents.
Workshop exercises used Python and SQLite, and some exercises used Datasette.

Total Codex token spend by workshop participants was 23 US dollars.
The workshop used GitHub Codespaces and OpenAI Codex to distribute a budget-restricted Codex API key to attendees for cost control and ease of setup.

A highlighted workflow configured Datasette to serve static content from a visualization folder, then used Claude Code to iteratively create interactive visualizations directly in that folder.
Claude Code produced a heat map visualization for a trees database using Leaflet and Leaflet.heat.

The handout was designed to be useful to people who did not attend the workshop, and the author expects it to apply beyond data journalism to anyone exploring data.

How many participants were in the workshop, and what was the distribution of token spend per attendee and per exercise?
What model versions, prompting patterns, and guardrails were used (e.g., constraints, system prompts, tool permissions) when using Claude Code and Codex?
What were the observed quality outcomes (correctness, data cleaning accuracy, visualization correctness, hallucination rates) versus a non-agent baseline?
What review and testing practices were used for agent-generated code and data transformations in the exercises?
How maintainable were the generated visualization artifacts over multiple iterations (dependency management, code structure, performance on larger datasets)?

Hands-on training using Claude Code and OpenAI Codex for data exploration, cleaning, and visualization suggests growing institutional interest in agent-assisted analytics workflows, which could translate into broader tool adoption across data-centric teams.
The described rollout pattern using Codespaces and budget-restricted API keys points to a governance-friendly path for deploying agent tooling in group settings, potentially increasing organizational willingness to trial paid LLM tooling.
Generating visualization artifacts directly into a served Datasette folder highlights demand for rapid agent-in-the-loop prototyping tied to deployable outputs, implying value for platforms that tightly integrate coding agents with lightweight data apps.

Evidence of sustained adoption of the handout beyond attendees, such as reuse in other trainings, public forks, or reports of teams applying the workflow outside data journalism.
Additional cohorts replicating the same low-friction setup, specifically standardized environments plus spend-limited keys, with consistent aggregate costs that remain manageable at larger group sizes.
Published evaluation of agent-generated code quality, including correctness of cleaning steps and visualization behavior, plus described review and testing practices that make the workflow reliable and maintainable.

Reports that agent-generated transformations or visualizations frequently require heavy rework, fail correctness checks, or become hard to maintain across iterations, limiting practical usefulness.
Observed token spend per attendee or per exercise materially exceeds expectations, making budget governance less effective for classroom or team rollouts.
Little to no follow-on usage outside the workshop context, indicating the workflow does not generalize well or lacks sufficient utility for broader audiences.