← Back to blog
AiAbout 6 min read

Adobe Is Turning Firefly Into the Agent That Runs the Apps

Published Oct 2, 2026
Adobe Is Turning Firefly Into the Agent That Runs the Apps

Adobe spent years treating generative AI as a feature bolted onto Photoshop. The version shipping now is closer to a layer that sits above the whole suite, which is a different product with a different set of problems.

The clearest expression of that shift is the Firefly AI Assistant. You open Firefly, pick the assistant from the left panel, upload your assets or references, and describe the outcome you want. Adobe then returns a set of suggested actions to refine or redirect, and the results can be saved to Creative Cloud storage. The company is packaging common multi-step jobs, such as Portrait Retouch and Social Media Assets, as prebuilt Creative Skills. The idea is that instead of knowing which of six apps to open and which tool inside each one, you say what you need and let the assistant assemble the path.

That path now reaches across Adobe's core apps. The Creative Agent moved into public beta across Premiere Pro, Photoshop, Illustrator, InDesign and Frame.io. For anyone who has watched a marketing team shuttle a single asset between three of those programs, the appeal is obvious. The friction Adobe is selling against is not generation quality. It is the handoffs.

Firefly Boards and the ideation problem

The most successful piece of this stack might not be the assistant at all. It might be Firefly Boards, Adobe's infinite collaborative canvas for moodboards, storyboards and in-place generation.

Boards is free with any Adobe account, and you only pay for premium models and higher limits through a Firefly or Creative Cloud plan. That pricing choice is what makes it interesting. It puts a generation surface in front of people who would never open Illustrator, and it lets them reach for models beyond Adobe's own, including Google's Gemini and GPT Image. In practice, Boards tends to get new models before Photoshop does.

The workflow logic behind it is sound. Early-stage creative work is messy. You are collecting references, testing directions, and throwing most of them away. A tool built for the final 10 percent of polish is badly suited to that phase. Boards is built for the first 10 percent, which is why designers reach for it before they know what the project is.

Adobe has been adding production features on top of that ideation surface. Presets turn entire generation workflows into one click, so a Restyle preset repaints an image in a new look and a Mashup preset fuses several images into one composition. A Rotate Object feature converts 2D images into 3D and lets you reposition people and objects to show different perspectives. Generative Text Edit changes the text inside a photo. PDF export and bulk image download smooth over the collaboration handoff.

Image Model 5 and the editing bet

At the model layer, Adobe's Firefly Image Model 5 generates photorealistic images at native 4MP resolution without upscaling, and it handles portraits with noticeably better anatomy than earlier versions.

The more consequential part is what the same model powers. Prompt to Edit lets creators describe an edit in ordinary language instead of brushing a mask. Layered Image Editing, still in development, aims at context-aware compositing that keeps a change coherent with everything around it. If that lands, it closes the gap between generating an image and actually working on one.

There is a catch already visible in Adobe's own documentation. Prompt to Edit supports premium models only, and users without premium-model access are told to generate a new image instead. An editing feature that quietly becomes a regeneration feature is a workflow difference, not a rounding error. It is the same complaint people have about every AI editor: the moment the tool cannot do the precise thing, it offers to redo the whole picture.

The data-routing footnote most reviews miss

Adobe routes a lot of Firefly traffic through partner models, and that has consequences the marketing rarely mentions.

When a partner model is used, Adobe's documentation says the prompt and any uploaded reference file are shared with that partner. Adobe states the content is never used to train partner or Adobe models, and that it only offers partner models where the developer or hosting platform contractually agrees to that condition. That is a reasonable position, and it is also a data governance decision every enterprise buyer now has to make on purpose.

The sharper detail concerns geography. Generating with Seedance 2.0, Kling 3.0 or Kling 3.0 Omni requires sending prompts and reference files to the Chinese companies that developed those models. Adobe cites contractual protections and technical safeguards, including the US DOJ Bulk Transfer Rule. These models are restricted to individual and Creative Cloud for individuals users. Adobe states plainly that partner models from China are not available through Creative Cloud for enterprise plans.

Most competing reviews get this backwards, treating model choice as a purely aesthetic matter. It is not. For a large organisation, picking a model is picking a data-handling regime, and the cheapest or newest model is not always the one whose paperwork clears legal.

The pricing trap nobody escapes

Adobe's credit model deserves a plain description, because it shapes how people actually use the tools. Firefly Custom Models, the feature brand teams want most, is priced at 500 credits per training run and 20 credits per generation, and the beta is tied to Creative Cloud plans rather than the standalone Firefly subscription some buyers assume covers it. Firefly Boards is free with any Adobe account, but generation still draws on the same credits and premium access as everywhere else, so an unlimited canvas is not unlimited output. There is also no published per-project or per-user consumption reporting, which leaves a team lead trying to attribute spend to a client without an official number to point at.

The bet underneath

Adobe is also previewing something called Project Moonlight, a conversational agentic interface that works across Adobe's apps and pulls context from a creator's own channels to suggest directions. Alongside it, Firefly Creative Production is expanding into a batch tool for editing thousands of images at once through a no-code interface.

Read the whole list together and the strategy is legible. Adobe is not trying to win the model race. It is trying to own the orchestration layer and the file formats, on the assumption that professionals will pay a premium for AI that lives where their work already lives. The public betas and gates are a way of testing that assumption without betting the franchise on it.

Whether it works depends on something unglamorous. If the assistant reliably saves a designer two hours on a real project, the subscription stays. If it takes three suggestions to get one usable result, the user goes back to doing it by hand and the agent becomes another tab they ignore. Adobe's own pricing and entitlement rules suggest the company knows exactly how thin that margin is.

Related articles