AI for Data Analysis I

This course

• You will use AI agents in your own research over the next five days.
• An agent is AI and it writes code, runs code on your data, and works inside your project.

What the agent did

• The agent wrote the code and ran it.
• It found 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.

More tasks to delegate

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

Your new role: direct the agent

• You already analyze data and use R, Stata, or Python.
• The new skill is directing an AI agent. It can write code and run that code.
• The agent made a country-level average in ninety seconds, but it is still your call if that average answers the question.
• You are the only person who can decide if a number is believable.

Who this course is for

• You wrote your own scripts and ran them. You loaded data and ran a regression.
• You provided a dataset or a current project, and every exercise uses it.
• You do not need a software engineering background.
• If you already use Claude Code or Cursor every day, this course will be too slow for you.

Pick one tool to use all week

• This course uses one coding agent. Claude Code (by Anthropic) works and Codex (by OpenAI) also works. Pick one and use it all week.
• The pre-course email explained how to install either one.
• Every student uses the tool's exact commands, and two instructors support a class much better when everyone runs the same setup.
• What you build moves regardless. The Skill format you learn on Wednesday is an open standard, and many agents read it.
• The agent runs on a computer, and you need a Mac or Windows laptop. Once it is running, you can use it from a phone or tablet.

Outline of the week

• Monday: Direct an agent to run your first analysis.
• Tuesday: Lead a complete analysis and at every stage, decide what to hand over.
• Wednesday: Teach the agent your standards, so every new session follows them.
• Thursday: Use the agent, collect data, and search the literature.
• Friday: Build a small tool for your own project.
• Every day builds your own project, and you leave with it on Friday.

Your Friday project

• You will present your own project by Friday.
• It does not need to be finished. Some will have results or a start.
• Your account is what matters. Tell your question, say what you built, and state how far you got.
• For each main step, say who did it: which parts you decided, and which parts the agent did.
• Tell us your experience and say what worked and what the agent did wrong and explain how you caught the error.

Today

• First, the concept: what an AI agent is.
• Then setup: get your tools running.
• Build your own research project after the break.
• We end with a check of your data.

Isn't this just ChatGPT?

• The same kind of model is underneath it, but the software built around it is different.
• Ask ChatGPT for the WVS analysis, and it gives you a block of code which you must run yourself.
• Give the agent the same request. It writes the code, runs the code, hits an error, and fixes the error. It repeats this until the chart exists.
• ChatGPT answers once and then stops, but the agent continues to work until the task is finished.
• That difference comes from three engineering problems, solved between 2023 and 2025. They are the three conditions below.

The agent works in one folder

• The biggest change from ChatGPT is where your work lives: in a folder on your computer.
• Pick one folder for your project and the agent works inside that folder.
• Your data, scripts, and notes are all inside together.
• The agent only touches items in that folder and does not touch the rest of your computer.
• The memory stays in the folder, and next time, it picks up where you left off and does not start over like a chat.

Condition 1: tool use

• A language model alone only makes text.
• Tool use gives it real actions. The basic set:
  • read a file — your data, your script, a codebook, a PDF;
  • write or edit a file — create a script, fix a line, or draft a note;
  • run a command — execute a script, install a package, or list a folder;
  • Search the web, read pages, look up documentation, and find a source.
• You can plug in more tools later, which include a browser or your reference manager, and you will use some of them this week.
• Every action runs on your computer, the result goes back into the model, and the model decides what to do next.

Condition 2: the agent loop

• A chatbot answers one time and an agent works in a loop until the goal is finished.
• On the WVS chart the loop ran like this:
  • it read the survey file and the codebook,
  • It wrote code, which computes country averages.
  • It ran the code and hit an error. The error was due to missing values.
  • It read the error and rewrote that step.
  • it ran again and produced the chart.
• The loop repeats, continues until the task is done, and also stops if you stop it.
• You control how often it asks you first. By default it asks before acting, and looser modes let routine actions through.
• You will use the strictest setting tomorrow.
• Whichever mode you choose, judging the result stays your job.

Condition 3: code execution

• Code tells a computer what to do. It can do any task and does not just do statistics.
• This is a statistical script, and it is one type of code. Code can also download data, collect and sort papers, clean text files, and build a website. This week covers mostly code that the agent writes.
• Code is just a guess until it runs. The agent runs its own code, reads the errors, fixes them, and runs the code again. This way, you get a result that works.
• Some tasks do not need any code, examples include reading a codebook or drafting a paragraph, and in those cases, the agent just reads and writes.
• The agent runs what it writes, which lets the agent finish a job, and it does not hand you code to run yourself.

The three conditions together

• Remove any one of the three and the system will fall back to a chatbot.

A short history

• ChatGPT: end of 2022.
• Tool use became reliable in mid-2023.
• The agent loop grew more advanced during 2023–2024.
• Code execution was dependable during 2024–2025.
• In 2025, the first agentic coding tools came out and combined all three in the terminal.
• Two years ago, AI coding tools might have disappointed you. They lacked one or two of the three conditions.

The agent is your RA

• You were the research assistant, and a professor gave you a task, and you went and did it.
• You are now on the other side and the agent is your RA.
• It is fast and takes instructions literally. A vague brief results in a vague result.
• Your job is the part the professor kept. You decide what to ask for and judge what comes back.

One warning for the week

• These models sound fluent, but fluent language can still be wrong.
• Fear the answer that is both confident and wrong.
• Check every result the agent gives you and do this every time.
• The rule for these five days: an unverified result is not a result.

Strengths and weaknesses

• A 2026 benchmark had AI agents run full social-science replications, and scored each step:

• Use it to write code and read results. Finding the data and the final judgment stay yours.

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

What is the terminal?

• The terminal is a text interface for your computer. You type a command and it prints the result.
• On a Mac, use the app Terminal, and on Windows, use PowerShell.
• Agents run in the terminal, and this is where programs run.
• You need very little. cd folder moves into a folder and ls lists what is there.
• One command matters today: cd into your project folder before you start the agent.
• The folder where you start the agent is the project it works in. Your files, its permissions, and its memory all stay there.
• Everything else you can ask the agent to type for you.

When something breaks

• Something will break today, and that is part of the course.
• The agent runs a command, and the command fails. It reads the error and fixes the error by itself, and you watch and steer.
• If you are stuck, tell the agent in plain words what happened.
• Error messages provide information and the agent reads them very well.

Daily essential commands

• Esc stops the agent while it is answering, and you can redirect it before it finishes.
• Esc Esc goes back to an earlier part of the chat and the agent forgets everything after that point.
• /compact asks the agent to summarize the current conversation, and the agent then continues from that summary. Use this when a long session slows down.
• /clear clears the chat and starts a new one in the same folder.
• /exit closes the agent. claude --resume reopens a past conversation and continues that conversation.
• On Wednesday, we explain how these commands affect the agent's memory.

Project Overview

• 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
  • Create the figures. Write the results.
• Do all of these on your own project this week.
• Start today at the beginning, get your data in, and take a first look.

Full project setup by agent

• You can ask the agent to create folders, ask it to write the first scripts, or also ask it to run them.
• It sets up projects well, and this is one of its first strengths.
• Many people do not know this and set everything up by hand.
• You say what you want and check what it made.

One project, one folder

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

• You have been adding to this one project all week.

Save points with git

• Git is the standard tool, and it keeps a history of a folder. Each save is a labeled point, which is called a commit, and you can go back to it.
• Ask the agent to commit before any big change and, if the change fails, ask it to return to the last commit.
• You do not need to memorize git commands, and the agent runs them, so only say commit, show the history, or go back.
• Today you must know these three things exist: save a point, list past points, and return to one.
• Tell the agent not to save your raw data in the history, which ensures the data stays in data/.

Why researchers keep a history

• You can always return to a working version.
• The history shows how the analysis changed, and your methods section must describe this.
• A collaborator can see all changes between two versions.
• You need three words today:
  • a repository is a folder with a tracked history
  • a commit is one labeled save point
  • .gitignore lists the files the history must never include
• You do not type any of this. Ask the agent to set it up and explain any part you want to understand.

What you have now

• Your tools are working and your project has its first script.
• Tomorrow you will run a full analysis in stages, and at each stage, you decide what to give the agent.

Your schedule for the rest of the week

• You do not write code but direct the agent, which writes and runs the code.
• You can give it every part of the work. It can find and collect data, clean data, build your measures, run the analysis, and draw the figures.
• You must decide what to ask for and also check what comes back.

Pick your project now

• Pick one project tonight and carry it through all five days.
• It could be a dataset for analysis, a question you delayed, or a messy file you did not have time to clean.
• Aim for something that looks good by Friday. This could be a clear figure, a short, readable report, or a small website that shows your result.

Exercise

• Work in pairs, three minutes each. Run your script on your own data while your neighbor watches, and cover three things:
  • your question, in one plain sentence
  • one mistake the agent made
  • how you found it
• The whole class follows and two or three volunteers show their work. An interesting failure is worth more than a finished chart here.
• Prepare for three questions:
  • What is your question, in simple words?
  • What error did the agent make, and how did you find it?
  • Where should this be by Friday?
• This is a rehearsal and Friday’s presentation asks the same questions. It also asks who did each part and if it was you or the agent.

Homework

• If your environment is still broken, pair up tomorrow. And get a fix slot with Bing before Session 2.
• Tonight, write one sentence and it should be the first analysis you want to run on your data.