Adobe EditAgent: AI Tool Routing for Multi-Step Photoshop-Style Edits

What happened

Adobe Research and the University of California, San Diego have built EditAgent, an AI system that edits images in the way a skilled Photoshop operator would, according to reporting by The Decoder.

Rather than sending every request to a single general-purpose editor and retrying until the output looks right, EditAgent selects a specialised tool for each step of an edit. A standard language model then picks the strongest result by weighing measured image data. The Decoder reports that the approach works, but that it needs about ten times the compute of the alternative.

The work is framed as a test of how well AI agents can operate professional creative software, a question that matters for any effort to have software agents carry out multi-step tasks on people's behalf.

Why it matters

The significance lies less in image editing than in the design pattern. Many AI tools are built around one model doing everything. EditAgent points towards routing work to specialised tools and using a general model to judge the outcome. For organisations deploying agents, the question shifts from "can the model do this?" to "which tool should handle each step, and how do we verify the result?"

The cost trade-off is equally instructive. A tenfold compute requirement means any gain in output quality has to be weighed against infrastructure spend and response times, particularly where agents operate at scale. Digital teams evaluating agentic workflows will need to measure quality and cost together rather than assume the more capable approach is automatically the right one.

The Renascence take

Most coverage will read this as a milestone for creative software. The more useful reading is about orchestration and verification, and about where a business should be willing to pay for extra effort. Image editing is a convenient test case because the quality bar is visible, and that visibility is exactly what many service agents lack.

The headline is about Photoshop, but the lesson is about trust. An agent that can choose its own tools and then judge its own output is only as reliable as the checks it applies, and the measured data is doing the work a customer would otherwise have to do by eye. Good experience design makes quality visible, so operators should ask where their equivalent of measured data comes from before handing an agent a multi-step task. If compute costs rise tenfold, the sensible move is to reserve that effort for moments where a visible error would genuinely cost the customer something.