Skip to content

Adobe Will Let Someone Else's Agent Open Your Photoshop File. Layers Are the Interface.

Adobe previewed local MCP integrations that let a third-party agent edit native Photoshop, Illustrator, Premiere and After Effects project files on your own machine, with layers and sequences still editable. Private beta is early 2027 and the command list is undecided. The part you control now is whether your file can be addressed at all.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A fanned stack of translucent vellum sheets on a dark walnut desk, each sheet holding one soft-edged shape, held by brass clips beside a scalpel
Editorial illustration about stacked, separable layers in a handoff; not a product screenshot or a model output.

Adobe used its MAX keynote on 10 November to preview something it has never offered before: an interface that lets an agent Adobe did not build make changes inside a Photoshop, Illustrator, Premiere or After Effects project file, on your computer, using the file's own layers and sequences.

The wording in Adobe's release is careful and worth reading twice. The integration makes changes directly in native project files on your computer, and layers, vectors, bins and sequences stay editable, so you can take over by hand at any point. What the release does not say is where the file contents and the prompts go, because that depends on the agent you point at the file and Adobe has not published those details. If the agent is also wired to other MCP servers, it can pull a brief from Google Docs, do the work in an Adobe app, and post the result to Slack.

That is the preview. The schedule is a private beta in early 2027, with no general availability date and no published list of which commands will be exposed. Adobe's vice president Deepa Subramaniam told TechCrunch that the company is still deciding where to draw the line, and that it has no desire to limit what the endpoint can expose. Those two sentences describe a product that is being defined while it ships.

Everything else in this article is my reading of what that means for the files you already have.

Two agents, one of them new

It helps to separate what shipped from what was shown, because Adobe's messaging runs them together.

Adobe's own Creative Agent is in beta now inside Premiere, Illustrator, InDesign, After Effects and Frame.io, and in the separate Photoshop beta app. It does the things an assistant does when it lives inside the application: apply an effect across every clip in a sequence, track a masked subject from frame to frame, adjust colour and exposure across a whole sequence, review a plan before it executes. Photoshop 28.0 shipped parameterised tools in the same spirit, including Relight with up to three movable light sources and Generative Restyle, which lands as its own layer with a slider.

None of that is the interesting part. The interesting part is that the caller becomes someone else's model. ChatGPT, Claude, Gemini or whatever you have running locally gets to open your file and change it. That moves the unit of collaboration from the application to the document.

Every unnamed layer is a guess

An agent working inside a PSD is not doing anything mystical. It reads the document structure, decides which object matches the instruction, and calls an operation on it. When the structure is unreadable, it guesses, and a language model guesses with complete confidence.

Which means the first requirement is not a better model. It is a file a stranger can navigate. The layers called Layer 12 copy 3, the adjustment baked into a merged raster, the eleven variants flattened into one smart object, all of that becomes the reason the agent removes the wrong thing and reports success.

You can test this today, without any beta access. Hand your working file to a designer who has never seen it, ask them to change one element and leave everything else untouched, and count how many questions they ask before they start. That count is your agent readiness. It is also your freelancer readiness, your holiday-cover readiness, and your readiness to reopen the project in eight months. The preparation work is not new. MCP just makes the cost of skipping it arrive faster and with a straight face.

A file organised as named groups, with adjustment layers kept as separate objects, and destructive steps kept out of the main stack, is not a file that looks prettier. It is a file that has been turned into instructions.

Editable is a promise about the format, not about your work

Adobe's phrase is that layers, vectors, bins and sequences remain editable. That is true and it is also narrow. It says the file format did not collapse into pixels. It does not say you can reverse what happened inside it.

An agent asked to "clean up the background" might delete four layers, relink a smart object, and rewrite a mask. Every one of those leaves the document editable. None of them leaves a trail unless you made one. There is no undo stack that survives the agent's session, and there is no diff view that shows you what it thought it was doing.

The rule that follows is boring and non-negotiable. An agent never opens the only copy. Snapshot before a run, keep the snapshot named with a date, and if the tool cannot produce a change log, build the diff yourself in two passes: compare composites for what changed in the picture, then walk the layer stack and the linked assets for what changed underneath. A deleted hidden layer or a relinked smart object can render identically today and break the next edit, so an identical composite is not a clean bill of health.

The operations Adobe shipped first are the tell

Look at the tools that went out first, before any of this agent business: Relight with three movable light sources, Generative Restyle as a layer with a slider, a Remove tool that now takes the reflection and the shadow along with the object, a Quick Compare toggle.

They share a property. Each one produces an editable result with parameters left on it. You can turn the effect down, move the light, remove the layer. That is not a coincidence, it is the design constraint that makes a generative operation usable by someone with taste, and it is the constraint I would hold any agent to as well.

So when you evaluate an agent that edits files, the question is not "did it do the thing". The question is "did it leave me a handle". An agent that flattens its decision into the composite has done work you cannot grade, only accept or throw away. An agent that leaves a named layer, a mask, and a parameterised effect has done work you can finish.

Adobe's own plan mode points the same way. You review and edit the plan before execution, which is a handle on intent rather than output. Worth using as a template for what you demand from the third-party version.

The cross-app chain is the part that actually changes production

The demo Adobe keeps describing is a brief in Google Docs, an edit in a Creative Cloud app, and a handoff in Slack, with one agent moving between them. That is a different kind of change from text-to-layer.

It means the brief becomes an input format. A one-paragraph brief with no asset list, no sizes, no approval owner and no naming convention is not something an agent can execute, and it is barely something a junior can execute either. The teams that get value out of this will be the ones whose briefs already look like instructions, because they had to be for human handoffs.

It also means the review gate has to exist before the agent runs, not after. If nobody owns approval, an agent that can touch every clip in a sequence will eventually ship something nobody approved, and it will look finished enough that nobody catches it. That failure mode does not need a malicious model. It needs a busy Friday.

What is honestly worth delegating

Renaming and organising layers. Relinking assets. Conforming a sequence to a new delivery spec. Applying one approved grade across forty shots. Pulling a shadow with an object. These are not creative decisions, they are the tax on creative decisions, and they are exactly the operations that are easiest to specify and easiest to verify.

What I would keep off the list for now: anything touching type, brand marks, or a client-facing comp, and anything where the correct answer depends on a judgement nobody has written down. Not because a model cannot do it, but because the verification cost exceeds the doing cost, and you will not notice the miss until the client does.

Where this leaves you before 2027

Nothing here is available to try, and the private beta will not include everyone who wants it. That gap is the useful part, because the file work is not blocked on Adobe.

Take the project you are most likely to hand off in the next six months. Rename the groups so a stranger could find the subject, the light, the texture and the type without asking. Move the destructive steps into their own saved states. Write the four lines of context that a competent freelancer would need: what the deliverable is, what must not change, what is still undecided, and who approves. Then run the stranger test.

If that file passes, you will be ready for the agent whenever it arrives, and you will have already collected the benefit from every human you hand it to in the meantime. If it does not pass, you have found the real blocker, and it was never the model.

Sources

Share this guide
X LinkedIn