When an AI Says “Done,” How Do We Know? My Summer at UCSB

This summer, I spent six weeks at UC Santa Barbara's Research Mentorship Program (RMP), exploring a question that seemed pretty straightforward at first: If an AI agent says it completed a task, how do we know it actually did?

It turned out to be a much more complicated question than I expected.

The Problem

AI agents are getting increasingly good at interacting with websites. They can search for information, click buttons, and complete tasks without someone guiding them through every step.

But just because an agent says it finished something doesn't necessarily mean it did.

That's what my research focused on: exploring ways to independently check whether an AI agent actually accomplished what it was supposed to.

A Research Side Quest... Then Another One

One of my favorite parts of the project was how many different areas of computer science it introduced me to.

I went into the summer expecting to work mainly with AI. Instead, I found myself learning about computer networking, computer vision, and other topics I hadn't really explored before.

Each field offered a different perspective on the same problem, and I enjoyed trying to connect them.

What made the research especially interesting was that there wasn't a clear set of instructions to follow. I'd start exploring one idea, come across something I didn't understand, and suddenly find myself spending hours learning about something completely different.

There were plenty of moments when an idea that sounded promising turned out to be much more complicated in practice. But figuring out why something didn't work, and what I could try instead, was often the most interesting part.

Presenting My Research

At the end of the program, I presented my work at the UCSB RMP Research Symposium.

After spending weeks thinking about the problem, explaining it to an audience that hadn't followed my research was a challenge of its own. It forced me to think about the bigger picture and why the problem mattered in the first place.

I also presented a research poster, which led to some interesting conversations and questions about the project.

More Than Just Research

Outside the lab, one of the best parts of RMP was being surrounded by students who were just as excited about their own projects.

Everyone was researching something different, and I loved hearing people explain topics I knew almost nothing about. Between research, exploring Santa Barbara, and spending time with friends, the six weeks went by surprisingly quickly.

By the end, I'd learned a lot, but I'd also made some great friends along the way.

What I Took Away

Before RMP, I imagined research as spending weeks becoming an expert in one very specific subject.

What surprised me was how often I ended up exploring completely different fields just to answer one question.

I started the summer wondering whether an AI agent could be trusted when it said "done." I finished it having explored topics I hadn't expected to study and with a much better understanding of why I enjoy research in the first place.

You can start with one simple question and end up learning about five completely different things while trying to answer it.