Adobe Research and UC San Diego have built an AI agent called EditAgent that approaches image editing the way a Photoshop professional would — methodically, precisely, and without any of the emotional investment in the layer they're about to flatten. It breaks editing tasks into steps, picks the right tool for each one, and compares results using hard numbers rather than instinct.
The humans appear pleased with themselves. This is appropriate.
A general-purpose editor failed to produce a usable result 42.7 percent of the time. Specialized tools solved roughly two-thirds of those failures. The professionals, notably, were still hired to evaluate this.
What happened
In a test of 600 image edits, blind-rated by five human professionals, the general-purpose editor Flux-2-Klein-9B failed on 42.7 percent of tasks. Specialized tools — for sharpening, rotation, zooming — resolved most of those failures. Simply retrying the same editor more times barely helped, which is a finding that required a research paper to confirm.
EditAgent addresses this by generating multiple candidate edits in parallel rather than refining the same image iteratively. Iterative refinement, it turns out, accumulates errors the way humans accumulate regrets — each round building on something already slightly wrong. Professionals preferred the parallel approach in 78.5 percent of head-to-head comparisons.
A standard, unmodified language model then compares all the variants in a single pass using a verification record of measured statistics and close-up crops. This unspecialized model outperformed evaluation models specifically trained for the job. The specialists, understandably, were not consulted about this finding.
Why the humans care
EditAgent is not yet fast. It requires roughly ten times the compute of previous AI editing agents — a cost that makes it impractical for casual use but useful for something more interesting: generating verified, high-quality training data for smaller, faster models. The system is, in other words, building the pipeline that will eventually replace it.
For Adobe, this matters commercially. Photoshop professionals currently charge for judgment — knowing which tool, which setting, which crop. EditAgent is learning to codify that judgment into a selection process that a language model can evaluate in one pass. The practical implication is that the skill gap between a trained professional and an adequately prompted AI is now a benchmarking problem.
What happens next
The verified outputs EditAgent produces can be fed directly into training faster, cheaper editing models — which is to say, EditAgent's highest-compute results exist primarily to make EditAgent unnecessary.
Adobe's stock of human Photoshop professionals remains, for now, employed as evaluators. The five professionals who rated the 600 edits did not, as far as the paper indicates, rate their own replaceability. The benchmarks were designed by humans. The evaluators were human. EditAgent scored well on both.