AI agents are not here to create — they are here to run the factory
There is an image that circulated for months across design and audiovisual production forums: a creative director staring at a screen filled with AI-generated variants, all technically correct, all editorially empty. The image captured something that productivity data could not: that the problem was never the speed of generation, but rather that no one had solved how to channel that speed toward a specific intention.
That is what is changing now, and the change arrives without fanfare. It arrives in the form of process architectures, reusable workflows, and integration protocols that transform generative models into systems with memory, judgment, and the capacity for self-correction. What was presented at the Upscale conference in San Francisco was not a demonstration of technical capabilities. It was, in a certain sense, the first draft of a new way of organising creative production at scale.
Magnific CEO Joaquín Cuenca Abela articulated it with surgical precision: the goal is not to generate impressive images, but to help people "show others what they have in their heads." That phrase, apparently modest, contains a complete reorganisation of the AI agent's role within a creative workflow. It is not the artist. It is the system that learns to interpret the artist with sufficient fidelity to reproduce them at scale.
The shift nobody wanted to name yet
During the first two years of mass adoption of generative tools, the debate organised itself around the wrong question: whether AI was going to replace creatives. The question was comfortable for media and for critics, but operationally irrelevant for marketing teams, content production units, or agencies with real deadlines. The concrete problem was not whether AI could generate an image, but that when it did, it generated something different every time — sometimes brilliant, sometimes a disaster — and in either case with no trace of the process that led to that result.
The most repeated complaint among creative directors working with these tools was not about technical quality. It was about reproducibility. You request a targeted modification and the model rebuilds the entire composition. You ask for stylistic consistency across assets within the same campaign and you get variations that share the colour palette only by accident. The output exists; the control does not.
What AI agents are resolving, in the version that companies like Magnific and Adobe are building, is precisely that deficit. They do not generate better. They generate within a workflow that can be audited, corrected, and replicated. Cuenca describes a generation of agents that work in loops: they generate, they review what they produced, they expose that process to the user, and they allow intervention at any point in the chain. The difference from the previous model is not the capability of the underlying model. It is the structure that contains it.
Adobe arrived at an analogous conclusion from its position as an incumbent. At Adobe MAX 2025 it presented AI assistants for Express, Firefly, and Photoshop described as conversational and agentic experiences that allow users to create and refine work through language within the tools themselves. It then accelerated with GenStudio toward what it internally calls an agentic content supply chain: a system that connects brand context, planning, creation, distribution, and reporting. This is not a new feature. It is a redesign of the entire content production workflow, with agents as operators of each stage.
WPP, from its position in global advertising, made its own bet in January 2026 with the launch of Agent Hub within WPP Open: an internal library of agents designed to package agency knowledge into reusable tools for clients. The logic is the same in all three cases: the value lies not in the model that generates, but in the system that directs it with accumulated institutional judgment.
What the Model Context Protocol does that interfaces cannot
There is a technical detail that is passing relatively unnoticed but that carries structural consequences: the Model Context Protocol (MCP). This open standard establishes secure, bidirectional connections between data sources and AI-powered tools, and it is being adopted by tools such as Anthropic's Claude Code and OpenAI's Codex to interact with creative platforms such as Magnific's workflows or Adobe's tools.
The operational impact is deeper than it appears. If creative tools become invocable from any AI interface compatible with this protocol, the entry point to creative work changes in nature. A designer could begin in a conversational interface, jump to a node-based visual workflow, return to a team collaboration space, and expose the completed process through an application programming interface. The creative suite ceases to be a collection of separate applications and becomes a production facility with shared machinery.
This carries market-power implications that deserve attention. For incumbents with broad suites, MCP is potentially a way to extend their ecosystem into spaces they do not directly control. For specialised startups, it is an opportunity to position themselves as interoperable tool layers without having to compete with the distribution reach of Adobe or WPP. The technical standard, in this case, reorganises who can become a relevant provider without having built a complete suite.
Gartner projects that 40% of enterprise applications will include AI agents specific to concrete tasks by the end of 2026, compared to less than 5% in 2025. McKinsey notes that the return comes from redesigning workflows, not from adding an agent as an accessory to existing processes. The technical warning is more interesting than the percentage: an organisation that grafts agents onto a dysfunctional workflow only accelerates the dysfunction.
Creative work and the question the market has not answered
With agents taking over more parts of the creative process — not just the outputs — real tensions emerge around employment in the creative economy. Brookings research on online independent work found that freelance workers in occupations most exposed to generative AI experienced a 2% fall in contracts and a 5% fall in income following the arrival of new AI tools in 2022. The World Economic Forum projects that 39% of relevant labour skills will have changed by 2030.
Those numbers do not say that creatives will disappear. They say that the skills that were valuable before are no longer valued in the same way, and that the skills now in demand are not the same ones. The risk that several executives present at Upscale named in different ways is identical: companies that treat creative agents as a headcount-reduction tool will discover too late that more assets produced without better judgment generate more noise and less impact. The trap is not technological. It is one of quality management at scale.
What does appear to be crystallising, at least among the organisations that have the access and resources to experiment, is a re-hierarchisation of creative skills. Netflix, Amazon, Apple, and other companies present at the conference are signalling that writing instructions for AI models is becoming a minimum entry point. The differentiating skill is workflow design: understanding how a concept moves from the brief through to references, assets, variants, approval, localisation, and distribution. The person who can encode that journey into a reusable workflow holds a position that models are not replacing, because it requires institutional knowledge, editorial judgment, and an understanding of internal approval processes that no generalist model possesses by default.
The short film "Candela" that the Magnific CEO presented at the conference as a production example was not attempting to demonstrate the technical quality of the outputs. It was attempting to demonstrate something else: that a specific creative vision, sustained by thousands of editorial curation decisions, can produce a result with genuine identity. The distinction matters because it points to the threshold where agents have utility and where they begin to require irreplaceable human direction.
Speed without judgment is not production — it is volume
The displacement this moment reveals is not about who creates, but about where value resides in the creative production chain. For decades, value was concentrated in the capacity for technical execution: the illustrator who mastered Photoshop, the editor who knew the Premiere shortcuts, the copywriter who produced ten variations in a day. That technical capacity is becoming a commodity. What does not commoditise at the same speed is the judgment about which variant is the correct one, why that colour communicates trust rather than coldness, how a brand behaves under narrative pressure, or why that scene cut destroys the emotional tension the sequence needed.
The clearest risk of the hasty adoption of creative agents is not that they produce poorly. It is that they produce well, fast, and without friction, and that this fluency conceals the absence of judgment until the brand damage is already visible. Organisations that flood channels with generated variants without a consolidated layer of editorial judgment will not be misusing the technology in a technical sense. They will be using a technology well for the wrong problem.
The structural value that agents create, when implemented with rigour, is not found in the generation of assets. It lies in making creative judgment replicable, audited, and scalable. That is the promise that differentiates a well-configured agent from a tool that simply produces more. And that difference, ultimately, is not defined by the model. It is defined by the organisation that decides what to encapsulate, what to approve, and what to discard.
