Testing AI coding agents is expensive. To see how well an AI agent performs, it has to work through hundreds of tasks step by step, and each step requires a call to a large language model. A single full test of a frontier agent can cost hundreds to thousands of dollars, and developers repeat these tests many times during development. SMU OUB Chair Professor of Computer Science and Vice Provost (Research) David Lo, collaborated with researchers from Shanghai Jiao Tong University (SJTU), has proposed EarlyEval, a method that watches an agent as it works and stops the run as soon as it is confident the agent will succeed or fail. On three widely used benchmarks, EarlyEval cut 13% to 26% of agent steps and up to 44% of tokens, while predicting the final outcome with 89% to 97% accuracy. The scores agents received changed by only one to two percentage points, and the overall leaderboard ranking stayed almost the same. Earlier cost-saving efforts mainly reduced the number of test tasks, while EarlyEval makes each task cheaper to run.