AI Agent Fixed 17 of 33 Broken Hackathon Builds, Median Fix Just 4 Lines

A developer who previously built all 164 submissions from a hackathon on isolated disposable computers — finding a quarter failed — then deployed a coding agent to attempt repairs on the 33 confirmed broken builds. Each frozen machine was rewound to its exact failure state, and the agent was given the failing command, its output, and strict instructions to make the smallest possible fix without altering tests or faking dependencies. A key design choice was that the agent could never grade its own work; the harness independently reran the build command and only counted an exit code of 0 as success. The agent fixed 17 of the 33 builds, left 15 unresolved, and one build hung indefinitely regardless of intervention. Most fixes were minimal, with a median change of just 4 source lines and 11 of the 17 fixes touching fewer than 10 lines.