Companies are giving AI agents real permissions faster than they are learning how those agents fail. So at Badger BuildFest 2026, we asked 27 UW–Madison students to do something most teams skip: try to break the AI agents they had just built, and show their evidence.
What they found surprised us. The biggest risks weren’t the AI model. They were in the checks, scorecards and guardrails built around it, the part a company owns and can fix.
The Exploration Brief is a new series from the Tech Exploration Lab that shares what we learn from student builders with the business leaders who need it. This first issue covers:
- Five patterns behind how student-built agents failed, from a wrong answer backed by a real citation to a safety check that blocked good work
- Five plain questions leaders can ask any team building or buying an agent, no code required
- The Art of the Break winners: My Magic UW, with honorable mentions for Proxier, Assay and MorrowLab
Read the Exploration Brief, Fall 2026 – Break It Before Your Customers Do (PDF)
We’ll be discussing these findings, and the topics for future briefs, with industry leaders at the Lab’s invitation-only Tech & Innovation Leadership Roundtable on November 19. If you’re experimenting with AI agents and would like to compare notes or be considered for an invitation, get in touch.
Congratulations to every team that took on the Art of the Break, and thank you to our judges, Sai Nagabhairava of Meta and Shivansh Gupta of the Tech Exploration Lab.