Once upon a time (like, six months ago), if you wanted to try five different chart types for the same dataset, you had to manually code each one. Copy-paste, tweak, run, squint, repeat. It was like cooking dinner by going to the supermarket five times.

Well, my friends, the kitchen has changed.

Code agents — tools like Claude Code, OpenAI Codex, or Claude’s chat with code execution — can now take a dataset, autonomously trydifferent visualizations, run statistical checks, and come back with insights. All while you sip your coffee (or wine, no judgment).

The “Try Everything” Workflow

Here’s the part that excites my inner (evil?) data scientist:

Step 1: Upload a CSV to Claude or fire up Claude Code with your dataset.

Step 2: Say: “Explore this dataset. Try different chart types to find the best way to show the main story.”

Step 3: The agent goes berserk (in a good way). It reads the schema, computes summary stats, tries a bar chart, then a slope graph, then a heatmap. It spots that one category is driving all the variance and writes: “This ridgeline plot best highlights the distribution differences.”

Step 4: You lean back and whisper: “Good agent.”

This is not science fiction. This is March 2026. Simon Willison recently ran a workshop at NICAR 2026 showing exactly this — and his key takeaway was that these tools are not just for developers.

Trust, but verify

Think of the agent as a chart speed-dater. It goes on 20 dates with your data and comes back saying “I think you should meet the slope graph — great chemistry.” But you still decide if it’s a match.

Ask yourself: Is the interpretation right? Could it be a data artifact? Is it actually surprising? The agent gets you to the “interesting chart” faster. Your brain still does the storytelling.

Remember my old post about flight tickets priced in yen instead of dollars? An agent would catch that before you build 15 histograms wondering why Tokyo-London costs $47.

So what?

If you work with data, try this: pick a dataset you know well, give it to a code agent, and ask it to explore and visualize. Compare its findings with what you already know. Be amazed (or horrified).

The point is (pun always intended) that code agents are not replacing your analytical brain — they’re giving it superpowers. Less time wrangling matplotlib, more time finding the story.

Now if you’ll excuse me, I have an agent exploring my Glassdoor dataset and I’m very curious about what it found…