Adobe is testing how well AI agents can use editing tools like a Photoshop professional would
Adobe Research, in collaboration with UC San Diego, has introduced EditAgent, an AI agent designed to handle image editing tasks by mimicking how a professional might navigate software like Photoshop. Instead of relying on one broad editing model applied in a trial-and-error loop, EditAgent breaks a task down and routes each step to the most appropriate specialized tool available.
The key architectural decision here is the use of a standard language model as an orchestration layer. Rather than judging output quality through a fixed scoring function, the language model evaluates candidate results by examining actual measured image data - pixel-level metrics and other quantifiable properties - and selects the best outcome from among the options produced. This keeps the selection process grounded in concrete evidence rather than subjective or heuristic criteria.
The agent-based approach reflects a broader shift in how AI systems are being designed for complex, multi-step tasks. In image editing, different operations - color correction, object removal, sharpening, compositing - have historically been handled by distinct algorithms, each tuned for a narrow purpose. A general-purpose diffusion model can approximate many of these, but often not as precisely as a dedicated tool. EditAgent's value lies in knowing which tool to reach for, and when.
The practical limitation is compute cost. EditAgent requires approximately ten times the computational resources of a more straightforward single-model editing pipeline. That gap makes deployment in consumer-facing products a challenge for now, though research benchmarks suggest the quality gains are real. Adobe has not announced any plans to integrate EditAgent directly into Photoshop or other Creative Cloud products, but the research signals where the company sees agentic AI fitting into professional creative workflows over the longer term.
