Coding Agents Can Solve the Task — But How Much Work Did They Waste?

While AI coding agents can successfully fix bugs, this achievement often conceals a great deal of wasted work like editing incorrect files or repeating commands. To measure this inefficiency, a novel experiment simulated an agent's decision-making across over 50,000 runs on real repositories with pre-defined bugs. This simulation method allows for a controlled analysis of wasted turns and unnecessary code changes, separate from the performance of any specific AI model. The ultimate goal is to determine if specific rules and workflows can guide agents to solve problems with less wasted effort, improving overall efficiency beyond a simple pass/fail result.