Adobe will open Photoshop and Premiere to ChatGPT, Gemini and Claude agents

Adobe will let customers run ChatGPT, Gemini and Claude agents directly against Photoshop, Illustrator, Premiere and After Effects, with a private beta of its Local Model Context Protocol (MCP) integrations planned for early 2027. The company previewed the integrations this week, and the stakes are straightforward: if creatives start working inside AI chat portals instead of Adobe's own interfaces, Adobe needs those portals to reach its tools — or risk losing the customer relationship it has spent decades building.

The technical promise is concrete. Outside agents will be able to edit native project files on customers' own computers while preserving layers, sequences and other editable elements. That detail matters because it means an agent working through ChatGPT or Gemini would not flatten a Photoshop file into a static image; it would operate on the live, structured document that a professional editor expects to reopen later.

What is Adobe actually committing to?

Adobe has not promised that every feature in its tools will be accessible to outside agents. Ely Greenfield, CTO and SVP of Creative Products at Adobe, was candid about the limits.

"I can't give you the absolutes right now. This stuff is still sort of well into development. I can give you the intent, which is not to block people or trick people or limit people into only working in certain ways," Greenfield said. He framed the MCP endpoints as a customer-retention play rather than a giveaway: "The nice thing about putting the MCP endpoints in our tools is, if you're using them, you have our tools, right? And you're still our customer. We're still providing you with value."

That is the commercial logic behind what Adobe is branding as open access. The company is betting that as creative work migrates toward AI agents and away from traditional menus and panels, the tools — and the subscription revenue attached to them — remain anchored in Adobe's software.

How does Adobe see the agentic future?

Greenfield described the strategy as two-track: Adobe builds its own agents into its tools, and it simultaneously pushes creative capabilities out to wherever customers happen to be working.

"I think the big message today is we are really leaning into that, both with the agents that we are bringing into our tools, but also with how we are bringing creative capabilities, technologies, tools, and workflows to wherever customers are doing their creative work. Whether that's in our tools, whether that's through interfaces they're going to build with Graph, or whether that's through the agentic portals, Gemini, ChatGPT, Claude," Greenfield said. "Our job has always been to power creatives in their workflows. We still will do that, even though the workflows will start to move out into other places, which is great."

He also connected agents to the broader shift toward so-called vibe coding — people building software by describing what they want rather than writing code. "And what agents are opening up through vibe coding is the ability for everybody to look at a computer the way a programmer looks at a computer, which is not just a tool to be used, but a tool for building tools," Greenfield said.

Why is creative judgement the hard part?

Greenfield drew a sharp line between a language model and an agent. The distinction rests on planning and self-correction.

"The big difference between an LLM and an agent is that an agent plans, but most importantly, an agent has a feedback loop. You give it a task, it makes a plan to go do it, and then it tries to accomplish it. It can then look and evaluate what it did and decide: did it do a good job, did it not do a good job, how do you go from there, right? And loop again," Greenfield said.

That feedback loop is exactly where creative work gets hard. Greenfield argued that judging whether content actually works for a specific brand or campaign is a skill most humans lack — and that Adobe's professional customers possess it in abundance. Handing that judgement to software is a different order of problem than generating an image.

"Understanding whether a piece of content is good or not for this particular use case or this brand campaign is a very hard task which most humans aren't good at. Our customers are all very good at doing this. Giving them an agent that they would trust to hand this off to, because the agent can make good decisions on its own, is a very high bar. That's the big challenge when it comes to making creative agents that are successful," Greenfield said.

He pointed to a practical lesson from Adobe's early plugin work inside external chat interfaces: typing is not always the best input method. "We're already seeing in some of the work we've done bringing our plugin into these other chats where I'm going to be chatting, but also sometimes it's much easier to point at something or to drag something than it is to describe it. Being able to pop up some interfaces inside of the agent becomes incredibly valuable," Greenfield said. Expect agents to blend conversation with visual controls, not replace them.

How does Adobe set its research agenda?

Greenfield described a tight, continuous loop between product teams and research leaders, in which domain experts and principal investigators work directly with the people shipping features.

"We have a much more evolved process by which our product teams work closely with our research leaders, who are the domain experts and the principal investigators. There's a very collaborative process by which our product teams inform the researchers about what problems matter to our customers and therefore to us. Our research leaders inform our customers about what is the bleeding edge, what's possible now, and what could be possible in the future," he said.

Planning runs on a rolling cadence, and the pace of change in AI has compressed it. "Like any product company, we have an iterative planning process by which we're always checking in, every quarter, month, whatever it is, on what our strategy is, what our roadmap looks like, and adjusting based on what's changed, which seems to be every day now," Greenfield said. The same discipline applies to research investment, balancing short-term bets against long-term and risky ones. His yardstick for a healthy lab is deliberately counterintuitive: "You definitely don't want every project to succeed, or you're not pushing hard enough."

Adobe also continues to train its own generative models, a decision Greenfield tied directly to market gaps rather than prestige. "We still build our own generative models in order to fill gaps in the market. Originally, the gap was people didn't trust these models. They didn't know what they were trained on. They thought they were going to be sued for it. We had to fill that gap," he said. Commercially safe, transparently trained models were Adobe's answer when legal fear dominated enterprise adoption of generative AI.

What does Adobe want agents to know about you?

The long-term differentiator, in Greenfield's telling, is context: an agent that knows a customer's creative style, brand, guidelines and company. Adobe's tools are positioned as the place where that context gets collected — through MCP connections or through Adobe's own interfaces.

"If you think about the value of an agent, it's both its ability to plan and evaluate, and its ability to collect and use that data on your behalf. Those are the things we're investing in," Greenfield said. As for competing with the major agent platforms head-on, he dismissed the idea: "The more the public agents, the common agents, get better at that, that's great. We'll focus on the stuff that they don't do."

The bet, in other words, is division of labor. OpenAI, Google and Anthropic can own the conversational front door; Adobe intends to own the creative context, the professional file formats and the judgement infrastructure behind it. The early 2027 private beta will show whether outside agents can actually deliver professional-grade edits — and whether Greenfield's stated intent not to "block people or trick people" survives contact with the features Adobe ultimately decides to keep behind its own walls.