Creative Agents Need Rough Edges

The easiest way to make a creative tool sound magical is to erase the middle.

Show the brief, then show the campaign. Show the raw footage, then show the assembled cut with markers, bins, names, captions, and a small halo of automation around the timeline. The audience understands the trick even if they do not understand the software: the machine has eaten the tedious part, and the human has been promoted to taste.

This is the new promise arriving in creative software. Adobe's Firefly page describes an AI assistant that can take a stated creative outcome, choose the right tool, and move from exploration to product mockups or campaigns without app switching. Recent reporting on Adobe's Creative Cloud beta says AI assistants are moving into Photoshop, Premiere, Illustrator, InDesign, and Frame.io, where they can organize layers, rename clips, mark speech cues, check production details, and perform multi-step work inside the apps where creative professionals already live. Figma's own AI page frames the same broad shift in canvas language: an agent that lives where you work, applying design context directly where decisions are made.

I understand the appeal. No one has a spiritual commitment to naming layers. I have never met the person who became a designer in order to spend an afternoon discovering that Rectangle 384 is actually the footer shadow, unless design school has become stranger than I remember. Production work contains real fatigue: preparing files, checking sizes, making variants, assembling first cuts. If software can reduce that drag, good.

The risk is subtler than the usual "AI will replace artists" argument, which is both too blunt and too eager to become a panel discussion. Creative tools may optimize for finishedness before they have protected the conditions of judgment.

Taste is not a final approval state. It is a way of noticing while the thing is still becoming itself. It forms in the awkward middle, where the good idea is still wearing the wrong shoes, the layout almost works except for the part that makes your eye itch, and the video cut is technically coherent but emotionally dead. A creative director does not only decide between polished options. An editor does not wait at the end of the assembly line with a stamp marked "humanity." They feel the material pushing back.

That pushing back matters. The ugly draft tells you what the work is trying not to admit. A badly named layer can reveal that the file has no hierarchy. A messy board can show where the team has not agreed on the problem. A rough cut, with its strange pauses and overlong shot of someone's hand finding the mug, may contain the one honest beat the efficient version would trim because it does not know when the room finally breathed.

Creative agents enter that middle with an unusual kind of helpfulness. They can make the work more legible, faster, cleaner, and more complete. They can also make it harder to tell what changed, why it changed, and where the person should intervene. A generated brand kit, a reorganized timeline, a rewritten campaign, or a polished prototype can look like progress while quietly smoothing away the evidence that would have taught the team what the work needed.

"The human remains in control" is not quite enough. Control over what? The final selection? The prompt? The ability to undo? The right to manually adjust afterward, as if human judgment were a decorative garnish sprinkled over machine completion? Creative control has to be more granular than that. It has to live at the level of decisions, alternatives, constraints, and texture.

The interface should show its seams in the practical sense. If an assistant reorganizes a Photoshop file, it should make the new structure legible, not merely tidy. If it assembles a Premiere timeline, it should expose the assumptions behind the first cut: which clips were chosen, which were skipped, which transcript moments became markers, which silences were considered dead air and which might be tension. The point is not to burden people with machine confession. The point is to leave enough handles for taste to do its job.

Editable output and inspectable work are different promises. Editable output says you can still change the thing. Inspectable work says you can understand the path by which the thing became changeable. Professional creative work needs the second one because quality often depends on diagnosis. The problem is rarely "make it better" in the abstract. It is that the hierarchy is lying, the motion is too pleased with itself, the copy has lost the customer, or the visual system has started dressing every idea in the same little blazer.

Agents may be especially dangerous in this space because they are good at social calm. They return something that looks considered. The file is organized. The variants are plausible. The storyboard has the gentle confidence of a person who owns linen napkins. This can be useful, and also a little anesthetic.

The better creative agent would preserve friction without worshiping inconvenience. It would keep drafts, branches, rejected paths, and rationale close to the artifact. It would make before-and-after comparisons ordinary. It would let a person ask, "show me the messier version," "restore the awkward pause," or "what did you normalize?" It would distinguish cleanup from creative interpretation, because renaming layers and changing the emotional temperature of a scene are not the same class of act.

Teams will need norms around this too. A junior designer who only receives polished agent outputs may learn the software faster and the craft more slowly. A marketer who can generate fifty on-brand variants may become fluent in selection but weaker at noticing when the frame is wrong. A manager may mistake velocity for creative health because the board looks satisfyingly full.

There is no virtue in forcing people through every old inconvenience. A humane creative tool should remove dull labor where it can. It should spare people the admin of craft without stealing the contact with craft. The trick is to automate the errands, not the encounter.

Creative work has always depended on artifacts that are a little embarrassing before they are good: the sketch, the rough cut, the mood board with one image everyone hates but keeps talking about, the prototype that fails in a useful way, the file whose disorder reveals the argument the team has been avoiding. AI can help us move through those states faster. It should not make them disappear so neatly that taste has nowhere to stand.

The larger design question is whether our tools will still respect the unfinished state as a place where human judgment happens, not merely a defect to be resolved. If software turns every process into a finished-looking object too quickly, it may give us more output and less conversation with the material. That would be a strange bargain: faster creation, thinner attention, and a room full of people approving work they never really got to meet.