The class started with a twist on a classic Brown intro-course programming assignment. For years, legendary computer science professor Andy Van Dam has asked CSCI 0150 students to code their own version of the video game Tetris, which normally takes more than a week to develop. AI can program it in seconds, but the Agentic Studio instructors added a new wrinkle: Turn the game upside-down. Instead of a game in which oddly shaped puzzle pieces are arranged as they fall down from the top of the screen, the students asked their agents to make a game where pieces ascend from the bottom.

“Now Claude starts making mistakes,” Fisler said. “Just flipping the game over caused things to get interesting.”

For Michael Pantano, a rising sophomore who took the class, the mistakes Claude made in creating upside-down Tetris cast the challenges involved in agentic programming into stark relief. While the code “kind of worked,” he said, it was full of small errors that made it clear that Claude Code didn’t quite understand the assignment.

“Claude just goes and makes the code you tell it to make,” Pantano said. “It doesn't think about the actual context of the project, so it ends up hallucinating a lot of the details, and the code just crumbles on itself.”

Fellow student Advay Bajaj noticed that the code Claude produced often worked well enough, but was poorly optimized for the task. That might be fine in the short term, but it becomes a problem when it comes time to expand the code base to handle additional tasks.

“There are often lots of different ways you can write a piece of code but only a few of those ways are actually going to be good,” Bajaj said. “The problem that arises is the code is not structured to support expanded functionality or future requirements. It ends up being so messy and convoluted that it becomes increasingly difficult for the developer to move forward and extend the codebase.”

Both students said that the class made clear that there will always be a role for competent software engineers. While AI agents may be able to take some of the drudgery out of coding by producing thousands of lines of code in seconds, software engineers need to understand the whole of the project to guide the process and assure the quality of the final product.

That becomes particularly clear during code reviews, the often terror-inducing sessions during which developers are asked to explain and defend their coding decisions. Code reviews are standard procedure in industry, and the Agentic Studio instructors made sure they were a key part of the course.

“I feel like code reviews for sure will be important in teaching with AI,” Bajaj said. “Students need to be accountable for their code, because it's very easy for a student to generate AI code for an assignment but not really know what it does. Code reviews and design crits are super important to make sure that the student is understanding and not just prompting AI to complete their assignments.”

That realization of the continuing human role in software development was a critical aspect, Littman said.

“We asked the students to explicitly defend themselves; what is it that you brought to this project?” Littman said. “I think through the process of doing that, students started to understand what role they were playing and what value they were adding, and I think they felt pretty good about that.”

‘What makes a good program’

Fisler, Krishnamurti and Littman said they’ll spend the rest of the summer digesting what they learned during the past semester before the department makes significant curricular adjustments. Fisler will offer a different experimental course for actual novices this fall. But one thing that emerged from the course, they say, is that agentic coding presents an opportunity to restructure introductory computer science classes in a more holistic way.

Writing code has always been just one part of the software development process. Students must also know how to clearly define tasks, understand the nature and scope of data and inputs, and devise tests that effectively weed out bugs. However, the instructors say, many students have long had to focus their energies on the coding part, often to the detriment of the larger picture. Perhaps now, they argue, is the time to reorient computing education in a way that shifts students’ attention to aspects of software development that are overlooked in traditional computer science education.

That’s certainly a lesson that Pantano took from his experience in Agentic Studio.

“I envision the future of computer science education being much more than just lower-level coding,” he said. “You really need to start with the fundamentals of what makes a good program.”

Krishnamurthi said the students’ willingness to join their professors in exploring a way forward bodes well for the future. Because it was experimental, the instructors wanted to keep the class small — approximately 20 students. But when they announced it, around 80 applied to take it.

“One thing that was clear to me is how amazing it was to do what we did at Brown,” he said. “There's the student culture of openness to exploration, which is why I think so many students were willing to sign up for this course even though we explicitly said, ‘We don't really know what we're doing here.’ The students basically said, ‘We’re trying an experiment? That’s awesome. Let’s go.’”