AI for Data Analysis I

This course

• Over the next five days you use AI agents in your own research.
• An agent is AI that writes and runs code on your data, inside your project.

What just happened

• The agent wrote the code and ran it.
• It hit errors, read them, and fixed them.
• It checked its own output.
• It returned a chart for you to inspect.
• Your only inputs were the instruction at the start and your reading of the chart at the end.

You can push it further

• The first instruction was only the beginning. You can continue:
  • "Validate the variable against the codebook."
  • "Handle the survey weights."
  • "Write a README so this re-runs with one command."
• For bigger jobs, you can split the work:
  • one agent validates the data
  • one writes the code
  • one drafts the report
• By Friday, you will have done all of these on your own project.

From tool user to director

• You already analyze data in R, Stata, or Python.
• The new skill is directing an AI agent to write and run that code.
• The agent produced a country-level average in ninety seconds, and whether that average answers the question is still your call.
• You remain the only one in the room who can say whether a number is credible.

Who this course is for

• You have written and run your own scripts: load data, run a regression.
• You brought a dataset or an ongoing project. Every exercise runs on it.
• You do not need a software engineering background.
• If you already use Claude Code or Cursor daily, this course will be too slow for you.

One tool, on purpose

• This course runs on one coding agent. Claude Code (by Anthropic) and Codex (by OpenAI) both work; pick one and use it all week.
• The pre-course email walked you through installing either one.
• Every hands-on uses that tool's exact commands, and two instructors support a room far better with one setup per person.
• What you build transfers anyway: the Skill format you learn on Wednesday is an open standard that many agents read.
• The agent runs on a computer, so you need a Mac or Windows laptop; once it is running you can also reach it from a phone or tablet.

The five days

• Monday: run your first analysis by directing an agent.
• Tuesday: direct a full analysis, deciding at each stage what to hand over.
• Wednesday: make your analysis re-run with a single command.
• Thursday: use the agent to search literature and draft writing.
• Friday: build a small tool around your own project.
• Each day adds to one project of your own, and on Friday you leave with it.

Where this is going: your Friday project

• By Friday you will present a project of your own.
• It does not need to be finished; some will have results, some will have a start.
• What matters is the account you give: your question, what you built, and how far you got.
• For each main step, say who did it: which parts you decided, and which parts the agent did.
• Share the experience: what worked, what the agent got wrong, and how you caught it.

Today

• First, the concept: what an AI agent is.
• Then setup: get your tools running.
• After the break, build your own research project.
• We close with a checkpoint on your own data.

Isn't this just ChatGPT?

• Underneath it is the same kind of model. The difference is the machine built around it.
• Ask ChatGPT for the WVS analysis, and it hands you a block of code to run yourself.
• Give the agent the same request, and it writes the code, runs it, hits an error, and fixes it, until the chart exists.
• ChatGPT answers once and stops. The agent keeps working until the task is done.
• That difference comes from three engineering breakthroughs, solved between 2023 and 2025.

It works in a folder, not a chat

• The biggest change from ChatGPT is where your work lives: in a folder on your computer.
• You pick one folder for your project, and the agent works inside it.
• Your data, your scripts, and your notes are all in there together.
• The agent only touches things in that folder, so it stays out of the rest of your computer.
• Its notes stay in the folder, so next time it picks up where you left off, instead of starting over like a chat.

Condition 1: tool use

• A language model on its own only produces text.
• Tool use gives it real actions: read a file, write a file, run a command, search the web.
• Every time it edits your script or runs python analysis.py, that is tool use at work.
• The model asks to run an action, the action runs on your computer, and the result goes back to the model.
• This is why the agent asks permission before it acts: it is running on your machine.

Condition 2: the agent loop

• A chatbot answers once. An agent works in a loop until the goal is met.
• On the WVS chart the loop ran like this:
  • it read the survey file and the codebook,
  • it wrote code to compute the country averages,
  • it ran the code and hit an error on missing values,
  • it read the error and rewrote that step,
  • it ran again and produced the chart.
• The loop repeats until the task is done, or until you stop it.
• It pauses for your approval, and that pause is where your judgment enters.

Condition 3: running the code

• Much of research work is code: cleaning data, running a model, drawing a chart.
• For that work, code you have written is only a guess until it runs.
• The agent runs its own code, reads what breaks, fixes it, and runs again, so you end up with a result that works.
• Some tasks are not code, like reading a codebook or drafting a paragraph, and there the agent just does the writing.
• Running its own code is what lets the agent finish the job.

The three conditions together

• Take away any one of the three, and the system falls back to a chatbot.

A short history

• ChatGPT: end of 2022.
• Tool use became reliable in mid-2023.
• The agent loop matured through 2023–2024.
• Code execution became dependable during 2024–2025.
• The first agentic coding tools, in 2025, combined all three in the terminal.
• If an AI coding tool underwhelmed you two years ago, it likely had one or two of the three, and this is why.

The agent is your RA

• You have been the research assistant: a professor hands you a task, and you go do it.
• Now you are on the other side, and the agent is your RA.
• It is fast, and it takes instructions literally, so a vague brief comes back as a vague result.
• Your job is the part the professor kept: decide what to ask for, and judge what comes back.

One warning for the week

• These models are built to sound fluent, and fluent language can still be wrong.
• The case to fear is the answer that is confident and wrong at once.
• So you check every result the agent gives you, every time.
• The rule for these five days: an unverified result is not a result.

What it is good and bad at

• A 2026 benchmark ran AI agents through a full social-science replication and scored each step:

• You lean on it for the code and the reading; you own the data and the final call.

Source: Nguyen et al., "ReplicatorBench: Benchmarking LLM Agents for Replicability in Social and Behavioral Sciences" (2026), arXiv:2602.11354.

What the terminal is

• The terminal is a text interface to your computer: you type a command, it prints the result.
• On a Mac it is the app called Terminal; on Windows it is PowerShell.
• Agents live here because this is where programs are run.
• You need very little of it: cd folder moves into a folder, ls lists what is there.
• Everything else you can ask the agent to type for you.

When something breaks

• Something will break today. That is part of the course, on purpose.
• When the agent runs a command that fails, it reads the error and fixes it on its own; you watch and steer.
• When you are stuck yourself, tell the agent in plain words what happened.
• Error messages are information, and the agent is very good at reading them.

What a whole project looks like

• Most research projects go through the same steps, and you can hand each one to the agent:
  • find and collect the data
  • clean it and get it ready
  • build the measures you need
  • run the analysis
  • make the figures and write up the result
• This week you do all of these on your own project.
• Today you start at the beginning: get your data in and take a first look.

The agent can set up the whole project

• You can ask the agent to make the folders, write the first scripts, and run them.
• Setting up a project is one of the first things it does well.
• Many people do not know this, and set everything up themselves by hand.
• You say what you want, and you check what it made.

One project, one folder

• Everything in this course lives in one folder for your own project.
• Ask the agent for a clear layout, so you and it always know where things are:

• You keep adding to this one project all week.

Where you are

• Your tools are running, and your own project has its first script.
• Tomorrow you run a full analysis in stages, deciding at each stage what to hand the agent.

How you work for the rest of the week

• You do not write code yourself. You direct the agent, and it writes and runs the code.
• You can hand it every part of the work: finding and collecting data, cleaning it, building your measures, running the analysis, and drawing the figures.
• Your job is to decide what to ask for and to check what comes back.

Pick your project now

• Choose one project tonight and carry it through all five days.
• It can be a dataset you want to analyze, a question you have been putting off, or a messy file you never had time to clean.
• By Friday, aim for something that also looks good: a clear figure, a short readable report, or a small website that shows your result.

Before you go, and tonight

• Show a neighbor your script running on your own data.
• If your environment is still broken, you pair up tomorrow and get a fix slot with Bing before Session 2.
• Tonight, write one sentence: the first analysis you want to run on your data.