Adobe launches a conversational AI assistant for Express and Photoshop, promising agentic creativity. Early demos reveal a system that executes commands mechanically but struggles with abstract direction, demanding relentless human oversight.
Adobe unveiled the AI Assistant for Express and Photoshop at MAX on October 28, 2025, pitching the update as a decisive move toward agentic workflow orchestration. The Express agent modifies layers, backgrounds, and typography via natural language while preserving untouched canvas regions. The Photoshop version entered private beta simultaneously. The gap between the autonomy Adobe markets and the correction operators must perform defines the product.
Prompt-driven iteration replaces judgment
The Express agent operates on command-response cycles. A directive such as "make this more tropical" triggers cascading adjustments to imagery, color palettes, and typographic pairings. The system processes the request mechanically, updating affected elements while isolating parts of the canvas marked as static. This preservation logic prevents total regeneration but highlights the agent's reliance on explicit scope definition.
Under the hood, the integration leverages the Model Context Protocol (MCP). External chatbots connect to the environment, and third-party plugins route through a dedicated Dev server. Adobe is structuring the application as a hub for agentic inputs rather than a siloed drawing board. The architecture assumes the model will receive structured goals and execute them reliably. Current demonstrations show reliable execution of stylistic shifts within defined bounds. Ambiguity outside those bounds forces the user to rewrite the prompt.
Governance limits full autonomy
Adobe's rollout reflects calculated risk management. The company enforces "Human-in-the-Loop" approval gates for design workflows, requiring human approval gates for design workflows. Deployed conversational agents operate under dual-tier human oversight frameworks. These controls acknowledge that unguided generation carries legal and reputational exposure. Enterprises cannot delegate creative authority to a black box that hallucinates trademarks or violates copyright filters.
Market sentiment skews positive despite the friction. An Adobe-commissioned survey of 16,000-plus creators recorded that 81 percent of creators had used generative AI experiences similar to Express. More than 70 percent expressed optimism about the creative upside of agentic AI. Creators anticipate value from speed and iteration, accepting that the initial results require editing. The demand exists for tools that reduce grunt work. The challenge lies in pushing the boundary of acceptable error rates downward.
Parallel developments suggest scaling complexity. "Project Moonlight" previews cross-application orchestration linking multiple creative applications. Connecting distinct media types increases the surface area for drift. If the agent misinterprets a style cue in a single image, propagating that interpretation across video timelines and photo sets multiplies the cleanup overhead. Orchestrating across apps rewards consistency. Inconsistency penalizes the operator harder.
Our read
The "mediocre intern" characterization holds because the agent performs tasks, not thinking. It adjusts pixels based on keyword association. It does not understand composition hierarchy, brand voice, or audience resonance. Professionals treating the tool as a collaborator will burn hours refining outputs that barely meet the bar. Treating it as a junior draftsman yields faster turnaround, but only if the operator retains full control of the creative brief.
We see a classic platform lock-in play masked as productivity. By standardizing on MCP and baking the agent into core applications, Adobe raises switching costs. Studios invested in this workflow face steep penalties moving elsewhere. The success metric won't be whether the agent writes copy perfectly; it will be whether the cumulative friction keeps teams locked inside the suite. The same structural trap awaits any studio that mistakes agentic UI for agentic intelligence, a dynamic explored in depth in AI Eats the Apprentice Tier.
Watch for Adobe to introduce confidence scoring or rejection thresholds that force earlier human review, shifting the burden from post-generation correction to pre-generation alignment.
