Image editing AI often handles simple requests well, but fails on complex ones, like moving a person, widening the frame, and changing the time of day in one go. Standard agents break the request into steps, but when a step fails, they simply retry with the same tools and same type of visual check.
EditAgent expands the agent's toolkit and the judge's evidence. It tries specialized methods in parallel and lets a vision-language model judge select the best result against a detailed report.
Professional editors preferred this method over traditional sequential refinement in 78.5 percent of comparisons. It used at most half the number of judgment calls. Because it only accepts verified edits, the training data it produces makes future standalone editors stronger.