Adobe expands AI agents and creative tools across Photoshop, Premiere, and Firefly

Adobe is expanding its AI strategy across Photoshop, Premiere, Illustrator, InDesign, Lightroom, and Firefly, pairing new in-app features with an agent system designed to take action across creative workflows. The company also announced a local Model Context Protocol (MCP) that can expose more Creative Cloud capabilities to external assistants such as ChatGPT, Claude, and Gemini.

The announcements, made at Adobe MAX in Miami, show Adobe addressing two pressures at once: the demand for faster AI-assisted production inside established creative applications and the rise of agent platforms that could become the main interface for software. The most direct competitive signal is Firefly Builder, a private-beta prototyping tool that takes Adobe further into territory associated with Figma.

Adobe broadens AI inside Creative Cloud

Adobe is adding targeted AI features rather than replacing the core workflows of its creative applications. Photoshop is receiving a relighting feature and controls for applying generative styles through prompts. Premiere is gaining a new professional color-grading system and an object-removal tool for video. Lightroom will offer tools to correct blinking eyes and remove blemishes.

The company is also making some actions easier to reach through what it describes as a more proactive, always-on agent. Adobe says the agent can use additional context about a user’s work and help with quick tasks, including background removal and video-thumbnail generation.

That approach matters for professional users because many common editing jobs are repetitive but still require context. Removing an object, adjusting a subject’s appearance, or preparing a thumbnail can be simple in isolation while becoming costly across a large content pipeline. The value of Adobe’s approach will therefore depend less on the existence of individual features than on how reliably they preserve the user’s intent and existing project structure.

Firefly moves toward prototyping and production

Adobe’s Firefly app is gaining Builder, a private-beta tool for creating prototypes that can communicate design ideas. The feature places Firefly closer to products such as Figma, although the available evidence does not establish how Builder compares with Figma on collaboration, code generation, version control, or production handoff.

Adobe is also introducing a tool for creating on-brand graphs and an image-editing function that can modify one part of an image without changing other objects. In video, a new Firefly tool will assemble projects from clips and additional elements such as images and sound effects.

These additions extend Firefly beyond standalone image generation. For product teams, the more consequential question is whether Adobe can connect ideation, brand controls, asset editing, and video assembly without creating a fragmented workflow. A prototype tool can be useful at the concept stage, but its practical impact will be determined by export options, review processes, editability, and compatibility with existing Creative Cloud files.

Adobe is also broadening the model choices available in its products. The company already offers models from Google, Runway, ElevenLabs, Black Forest Labs, OpenAI, Kling, and ByteDance. Topaz Labs, one of Adobe’s newer acquisitions, is releasing Wonder 4 for upscaling and denoising and ImageHDR for converting standard images into HDR.

MCP puts external assistants closer to Adobe workflows

Adobe’s local MCP is designed to let other chatbots call a wider range of functions in Creative Cloud applications. Adobe already has integrations with ChatGPT and Claude for selected product actions; the MCP is intended to provide a broader interface across applications including Illustrator, Photoshop, After Effects, and Premiere.

Adobe vice president Deepa Subramaniam told TechCrunch that the company is still determining which capabilities to expose and will use feedback from users to guide that process. Her comments indicate that the MCP is an evolving access layer rather than a finished catalog of every Adobe operation.

That distinction is important for developers and enterprise buyers. An MCP can reduce friction between an assistant and creative software, but it also raises questions about permissions, auditability, destructive actions, data handling, and the reliability of multi-step tasks. Adobe’s decision to release the MCP locally may offer a more controlled deployment model, but the source evidence does not provide detailed security or governance specifications.

The announcement also arrives amid debate over MCP itself. TechCrunch reported that Figma co-founder Dylan Field had called the standard imperfect and argued that it benefits frontier AI labs more. That criticism highlights a broader industry question: whether interoperability standards will give software users flexible agent access or primarily strengthen the companies operating the most capable assistants.

What the announcement means for builders and enterprises

For creative teams, Adobe’s release creates a larger set of automation possibilities inside tools they may already use. A production workflow could combine an external assistant with Photoshop edits, Premiere actions, and Illustrator operations instead of requiring a user to switch manually between an AI chatbot and Creative Cloud.

However, adoption will depend on controls as much as convenience. Enterprises will need to understand which assets an agent can access, whether actions require confirmation, how changes are recorded, and whether an automated edit can be reversed. Brand-sensitive teams will also need to test whether on-brand generation and graph creation produce consistent results across campaigns rather than merely attractive individual outputs.

For builders, the MCP may be more significant than any single feature because it creates a possible integration surface around Adobe’s installed base. Yet an open-ended endpoint strategy can be difficult to productize until the available actions, authentication model, rate limits, and failure behavior are documented. Adobe’s stated reliance on user feedback suggests those details may change as real workflows emerge.

The performance and adoption picture remains limited. The source material reports product announcements and Adobe’s own plans, but it does not provide independent benchmarks, usage data, or customer evidence for the new tools. Claims about productivity should therefore be treated as unverified until users test them in production settings.

What to watch next

The first signal will be the scope and stability of Firefly Builder as it moves through private beta. Its usefulness will depend on whether prototypes can move cleanly into design review, collaboration, and implementation.

The second is the practical documentation around Adobe’s MCP. Developers will be watching for the exact endpoints, permission controls, local deployment requirements, and support for multi-step actions across Creative Cloud applications.

Adobe’s model strategy is another key indicator. Supporting models from multiple providers gives users more choice, but it may also create differences in quality, cost, licensing, and behavior inside the same workflow. Finally, enterprise buyers will want independent evidence on reliability, asset protection, and reversibility before allowing always-on agents to operate across production files.

Creati.ai perspective

Adobe is not simply adding generative features to individual apps; it is positioning Creative Cloud as an operating layer for AI-assisted creative work. Firefly Builder addresses the front end of ideation, while the MCP addresses how agents may reach the tools where finished assets are produced.

The opportunity is substantial, but the hard part will be operational trust. Adobe’s announcements establish a wider set of capabilities, not yet a proven autonomous workflow. For builders and enterprise teams, the next phase will be less about counting AI buttons and more about testing permissions, edit fidelity, interoperability, and measurable time saved.