25. Lab: Building a Data Workflow with an AI Agent

In Working with GenAI Tools you learned how coding agents work: a harness runs a model in a loop, the model reads files like AGENTS.md, edits code, runs tests and acts on the results. In lectures 22 to 24 you designed a workflow to calculate a time series of mean NDVI from Landsat for two NEON sites, SJER and HARV. In this lab you bring the two together and build that workflow with an AI coding agent (GitHub Copilot in Agent mode).

The lab is built around one idea: when code is cheap to write, your value lies in the parts an agent cannot do for you. You decide what question to answer and which data fit it, you write the plan, you specify the tests that define what correct means and check that they can fail, you review every change, and you check that the results make scientific sense. The agent writes most of the implementation.

25.1. Who does what

Step	You	The agent
Question, data and outputs (lecture 22)	Write them in AGENTS.md	Reads them in every session
Plan (lecture 23)	Write the pseudocode, with inputs and outputs	Critiques the plan before any code exists
Tests (lecture 24)	Write the test plan with expected answers worked out by hand; review the tests; break the code on purpose to check they can fail	Writes the test code from your plan, then runs the tests and works until they pass
Implementation	Review every change and commit it	Writes the code, one step at a time
Results	Check them against scientific expectations
Code review	Find the problems in code an agent wrote without guidance

25.2. Getting started

The lab lives in its own repository: ndvi-workflow-lab. Click “Use this template” to create your own private copy, clone it, and follow README.md. You need git, pixi, VS Code and GitHub Copilot (free with the GitHub Student Developer Pack).

pixi install
pixi run test

25.3. Parts

Each part has its own instructions in the prompts/ folder of the repository.

Part	What you do	Time	AI?
0	Set up the repository and check that the tests run	10 min	No
1	Write AGENTS.md: question, data, outputs and workflow	25 min	No
2	Ask the agent to critique your plan, then update it	15 min	Yes, as a reviewer
3	Write a test plan for one scene, then have the agent write the tests	30 min	Plan: no. Test code: yes
4	Build the one-scene functions with the agent, then break them on purpose to check the tests	25 min	Yes
5	Build one site, then many sites, and run the workflow	30 min	Yes
6	Check the results against what you know about the two sites	10 min	No
7	Review a script an agent wrote from a one-line prompt	20 min	No
8	Reflect on what worked and what did not	homework	No

25.4. Habits to practice

• Commit before every agent session. If a session goes badly, pixi run reset returns you to your last commit.

• Use a new chat session for each part. Short, focused sessions are easier to review and use fewer credits.

• Read every change with git diff before you commit it. If you cannot explain a line, ask the agent to explain it, or change it.

• Never let the agent change your tests to make them pass. If a test is wrong, you fix it.

• Make sure every test can fail. A test whose expected value comes from the same formula as the code checks nothing.

• Passing tests show that the code does what you specified, not that the result is right. Always check the results against what you expect from the science.

• Keep notes as you go in ai_logs/log.md, written by you.

25.5. What to submit

Push these to your copy of the repository: AGENTS.md, your test plan (tests/TEST_PLAN.md) and tests, your code in ndvi_workflow/, the csv file and plot in outputs/, ai_logs/log.md, REVIEW.md and REFLECTION.md.