Assignment 1: Explore an Agent in a Domain You Know

Due: Monday, September 14, 2026, 12:00 noon.

Modern AI systems are increasingly agentic. They don't continue a conversation or answer a question, they take on a task and search, compare, plan and decide for themselves what to do next. You may have used this capability already, but in this project you will explore how well an agent can accomplish real task from a domain you know well.

Choose a small or task from your own field, work, or life that an agent can try to address by doing something. This could be something like figuring out the update rate of a webcam, finding appointment availability across multiple clinic websites, measuring how often a public data feed updates, or tracing a quote or image to its original source. The task should be one where you know enough to judge whether the agent's answer is right.

Run the task with an agentic tool of your choice. Then write a short blog post ("three screenfulls") that includes:

1. The domain. The task. Why does it matter to you, and what do you know about it that the AI doesn't?
2. The task. The exact prompt you gave the agent, verbatim, and what a correct answer would look like.
3. What the agent did. Walk through the steps it actually took, with a screenshot or transcript excerpt or two. Where did it surprise you?
4. How you know whether it's right. Did the agent check its own answer? Was that check real? Could you verify the answer independently — and did you?
5. Would you delegate this again? If the agent ran this task next week without you watching, what could go wrong, and what would need to be true for you to trust it?

An agent that fails at your task, examined carefully, makes just as good a blog post as one that succeeds. The interesting question is never only "did it work" but "how would anyone know."

The task should need under about 30 minutes of agent time.

Getting agent access for free

You do not need a paid account for this project. Options, roughly in order of ease:

• ChatGPT free tier (chatgpt.com). Free account includes 50 agent (Codex) messages per day in the web app. This is the simplest route and requires no setup.
• GitHub Copilot Student (education.github.com) verify with your GWU email to get the free student plan, which includes agent mode with a monthly credit allowance. Verification takes hours to a few days, so start early; a few long agent sessions can exhaust the monthly credits
• Fully local Ollama (ollama.com) runs open models on your own laptop at zero cost; suitable if your task involves data you don't want to send to a provider. Expect weaker agent behavior from small local models, which is itself worth writing about.

These free tiers change frequently; if one is unavailable when you start, use another and note it in your post.

Submitting your project

Submit a link to your blog post here: Assignment 1 submission form.