Study finds coding agents waste up to 22.75% of task costs

Researchers studied coding agents working on SWE-bench tasks and found three recurring ways they use resources inefficiently: retrieving information they already have, generating similar scripts, and rerunning tests. These behaviors appeared in 79% to 98% of tasks and accounted for as much as 22.75% of task costs. The team tested several ways to reduce the waste, including guidance written by developers. That approach cut costs by up to 41.73%; automatically generated guidance was less effective, while one retrieval method sometimes increased costs.