The end of the prompt era: AI that produces entire campaigns is coming

The illustrations in this article were generated by AI (Nano Banana Pro model on the FOTOhub.app platform). Labelled in accordance with Article 50 of the AI Act.

The most important skill in generative AI is no longer writing the perfect command. Give the system an objective, a budget, brand context, and operating boundaries, and it can map out the work, select the tools, produce variants, run tests, and return with the results. The prompt is not disappearing. It is simply ceasing to be the product.

On September 10, 2026, OpenAI released the public beta of its Agents API. It provides infrastructure for running agents that can operate over extended periods, preserve context, use tools, execute code, and delegate tasks to specialized subagents. The developer defines the task, model, tools, and environment; OpenAI maintains the execution layer that coordinates the process.

That launch alone will not automatically produce an advertising campaign. On September 10, OpenAI did not announce "an autonomous marketing agency in a single API." Claiming otherwise would be misleading. What the company did release, however, is a missing piece of the infrastructure: persistent sessions, context management, tool discovery, parallel tool calls, MCP support, subagents, and environments in which an agent can work with files and produce artifacts.

At the same time, Adobe, Salesforce, Google, and WPP are moving agents directly into marketing workflows. Adobe is developing an agentic content supply chain spanning briefing, creation, approval, activation, and measurement. Salesforce allows marketers to define objectives, budgets, constraints, and levels of autonomy, while an agent builds, launches, and optimizes a campaign within those boundaries - although some of those capabilities remain in pilot. Google has announced Ask Advisor, which connects agents with Google Ads, Analytics, and Merchant Center, while WPP Open brings strategy, creative, media, and production into one system.

These are not four unrelated launches. They are four versions of the same answer to a problem generators never solved.

A generator creates an asset. An agent manages an outcome.

A generator waits for a prompt. An agent starts with an objective.

A generator returns an image, a piece of copy, or a video clip. An agent should understand why that asset exists, who it is intended for, who must approve it, how much it may cost, and what to do if the first version fails.

This is why what is ending is not so much the era of prompts as the era of the prompt at the center of the product. We are moving toward a system in which people will no longer manually produce every individual component of a campaign. They will define direction, boundaries, risk, and success criteria, while AI performs an increasingly large share of the work between intent and outcome.

The Prompt Is Not Disappearing. Its Role Is Changing

"The end of the prompt era" must be understood precisely. Models will still require instructions. Agents also operate on objectives, context, and rules. We will not stop telling systems what we expect from them.

What will disappear is the need to treat prompting as the manual programming of every individual operation.

During the first phase of generative AI, users had to learn the language of the machine. They described the framing, lighting, lens, camera movement, style, composition, tone of voice, copy length, and response format. If the result was poor, they rewrote the prompt. If they needed a video based on an image, they moved the file into another tool and wrote another instruction. Voice, music, subtitles, resizing, and publishing all happened in subsequent systems.

That was progress compared with traditional production, but it was not autonomy. The human still acted as the manual router for the entire process.

An agent reverses that relationship. Instead of receiving a command such as "generate a product image on a bright background, then prepare a vertical video," it receives a business-level assignment: "Build a launch campaign for a new product across three audience segments, within a defined budget, in accordance with the brand guidelines, and do not publish anything without approval." It must then decompose that objective into steps, select data sources, tools, and models, manage dependencies, evaluate intermediate results, and stop wherever a human decision is required.

OpenAI defines an agent as a system equipped with instructions, constraints, and tools that allow it to act on a user's behalf. In its practical guide, the company reduces the core architecture to a model, tools, and instructions, while also emphasizing layered safeguards, access control, and human intervention for sensitive or difficult-to-reverse actions.

Anthropic draws an equally important distinction. A workflow guides models and tools through a path predefined in code. An agent determines its own next steps and decides how to use its tools. The company recommends starting with the simplest viable solution because a predictable workflow will often outperform an autonomous agent in well-defined tasks.

That distinction is critical in marketing. Automatically resizing an approved banner into five formats does not require a "thinking team of agents." It is a straightforward, deterministic process. Deciding which segments require separate creative, which variants should be produced, which model best suits the product, and when weak performance should be escalated may require a decision-making layer.

A mature platform will therefore not deploy an agent everywhere. It will use autonomy only where autonomy makes economic sense.

A Campaign Becomes a Program

A traditional campaign is a project. It has a brief, a schedule, a team, a list of formats, revision rounds, a media plan, and a final report. Every stage triggers a different specialist or application. The strategist hands the work to a copywriter, the copywriter to a designer, the designer to a motion specialist, production localizes the assets, legal reviews the risks, the media buyer publishes the campaign, and the analyst evaluates the results.

An agentic campaign behaves more like a program that remains active.

It has a primary objective, inputs, policies, a budget, permissions, triggering events, tools, approval thresholds, and stopping conditions. It does not end when a file is exported. It monitors signals, detects change, commissions new variants, and routes them for approval.

The biggest change, therefore, is not that AI will write a headline and generate a video. We can already do that. The change is that those components are being connected into an operational system capable of maintaining the objective throughout the process.

The OpenAI Agents API demonstrates why execution infrastructure is becoming a product category in its own right. An agent can operate inside a persistent session, while older context is automatically compacted to preserve the information required to continue the task. Tools can be loaded on demand, operations can run in parallel, and subtasks can be delegated to subagents with their own contexts.

For a campaign, this could mean separate roles for a strategy agent, a brand-research agent, a creative agent, a legal-validation agent, a variant-production agent, and an analytics agent. It does not, however, mean that more agents automatically produce a better system. Every handoff creates cost, latency, and a new point of failure. Anthropic recommends simplicity, transparent planning, and rigorous testing of tool interfaces, and argues that a multi-agent architecture should be introduced only when a simpler design is no longer sufficient.

Major Players Have Moved Beyond Generation

The strongest signal does not come from a presentation about the future. It comes from products that the industry's largest companies are trying to place inside the core marketing workflow.

Adobe Is Building an Agentic Content Supply Chain

Adobe Brand Intelligence is designed to transform static brand guidelines into context agents can use. According to Adobe, the system is intended to learn not only from documents but also from comments, rejections, and approvals, allowing future assets to align more closely with an evolving brand identity. The company connects this with Workfront agents, brief generation, production automation, review and approval workflows, and delivery of final assets.

In GenStudio for Performance Marketing, Adobe has made Content Production Agent available in beta to licensed customers. The agent receives a brief, objectives, and channel requirements; proposes a direction, personas, products, templates, and assets; and then generates variants in line with brand guidelines, while leaving users able to edit the output and submit it for approval.

This matters: even Adobe does not describe the current process as unrestricted autonomy. The product is embedded in existing assets, templates, brand policies, and approvals. The agent is not replacing the chain of accountability. It is intended to accelerate it.

Salesforce Wants to Manage the Objective

Salesforce has presented a model in which the marketer defines goals, budget, guardrails, and autonomy limits, while the agent builds an audience, launches a campaign, tests messaging, selects channels, and adapts the offer to customer behavior. This is precisely the transition from managing individual steps to managing an outcome.

The product status must, however, be stated honestly. In the company's June 2026 announcement, Agentforce Content Agent and Marketing Goals Agent were in pilot, while other components, including selected prospecting agents, were generally available. Salesforce itself warns that some of the services it describes may remain under development.

That distinction is critical. The agentic marketing market is moving faster than production deployments. The vision is real. Fully autonomous campaign execution that is reliable, widely available, and genuinely end to end is not yet the standard.

Google Is Connecting Objectives to Channel Data

Google Ask Advisor is intended to coordinate agents across Google Ads, Analytics, Google Marketing Platform, and Merchant Center. In one example, a user sets the objective of finding new customers for hair-care products, and the system retrieves product information from Merchant Center and helps configure a campaign in Google Ads. Ask Advisor is also intended to combine results from Ads and Analytics, explain what worked, and recommend the next actions.

Google stated that the solution remained in beta for English-language accounts, with broader availability expected later. This is further evidence of the direction of travel, but it does not justify claiming that Google has already given every advertiser a fully autonomous media buyer.

In parallel, the company launched a global beta of text guidelines for AI Max and Performance Max. Advertisers can use natural language to specify terms and concepts the system should avoid. It may look like a small interface detail, but it signals a significant architectural shift: as automation expands, so does the need for explicit brand boundaries.

WPP Is Encoding Agency Expertise

WPP Open connects strategy, creative, media, and production and, together with Adobe, is intended to orchestrate the planning, creation, activation, and optimization of marketing assets. In this arrangement, Adobe supports content production and adaptation, while WPP agents can optimize media and activate campaigns across channels.

WPP's Agent Hub demonstrates another dimension of the transition: an agent can represent not only an executor but also an organization's codified expertise. WPP has described agents grounded in brand data, behavioral science, and its own strategic practices. The company reported that WPP Open had more than 75,000 users across the organization and was used by more than 90% of employees working with clients.

If this layer proves genuinely useful, an agency will no longer sell only access to experts' time. It will sell access to their expertise embedded in processes, data, and agents.

A Campaign Agent Is Not a Single Model

The mistake of the first wave of generative AI was to confuse the product with the model. The same mistake is now returning under the label "agent." Part of the market is attempting to present a standard chatbot with a handful of functions as an autonomous marketing system.

A real campaign agent requires at least eight layers.

Objectives and Success Criteria

"Create a campaign" is not an operational objective. The system must know the market, audience, product, timeframe, budget, channels, metrics, prohibited actions, and hierarchy of priorities. It must also distinguish an intermediate objective, such as click volume, from the actual business objective.

Brand Context

A logo and color palette are not enough. Brand context includes language, positioning, approved claims, legal risks, examples of accepted and rejected creative, image-rights policies, local restrictions, and the brand's relationship with a specific customer group.

Adobe Brand Intelligence is moving in this direction by turning feedback from review and approval processes into part of the brand context available to machines.

Campaign Memory

The agent must know what it has already generated, which versions were rejected, which messages should not be repeated, and which decisions must remain consistent across channels. A long conversation alone is not a reliable production memory. The system needs a controlled project state, versioning, and a clear separation between durable data and the model's temporary context.

A Registry of Tools and Models

The agent should know which engine generates images, which produces video, which refines copy, which synthesizes voice, which removes backgrounds, and which publishes the result. It must understand parameters, pricing, latency, availability, usage rights, and data restrictions.

No single model is best at every stage. The agentic layer does not eliminate multi-model architecture. It gives that architecture a purpose.

Orchestration

Some steps can run in parallel; others can begin only after the preceding stage has been approved. One creative asset may require regeneration, while another needs only resizing. The orchestrator must manage dependencies, retries, timeouts, alternative providers, and the state of a partially completed campaign.

Evaluation

An agent cannot declare an asset successful simply because it generated it. It needs criteria covering alignment with the brief, brand, platform policies, channel format, legal requirements, and technical quality. The evaluator - optimizer pattern - in which one component creates and another assesses the result and requires improvements - is one of the foundational architectures described by Anthropic.

Approvals and Permissions

Generating a draft is easily reversible. Spending budget, sending a message, publishing an advertisement, or modifying a live campaign is not. OpenAI recommends human intervention for high-risk actions and after failure thresholds have been exceeded. Its approvals documentation describes a mechanism through which tool execution is paused and the same process can be resumed after a human decision.

Observability and Auditability

An organization should know which agent made a decision, what data it used, which model it called, how much the operation cost, who approved publication, and how the outcome shaped the next action. OpenAI's Agents SDK records items such as generations, tool calls, handoffs, and guardrails in its traces, allowing teams to analyze not only the output but also the process that produced it.

Without these layers, there is no autonomous campaign. There is only a demo.

Context, Not the Prompt, Is the Real Currency

A prompt can be copied. A model can be replaced. Inference costs are falling, and access to new engines is becoming increasingly commoditized.

It is much harder to copy a well-structured brand memory, the history of decisions, creative-performance data, approval rules, legal policies, and information about how assets move through a real workflow.

That is why the competitive advantage is shifting from prompt engineering to context engineering. Anthropic defines the term as selecting and maintaining the optimal set of information available to the model during inference. In a campaign agent, context extends far beyond a system instruction: it includes product data, assets, audience segments, behavioral signals, test history, tools, permissions, and the current state of the process.

This changes how AI products must be designed. The text field will remain useful, but the most important operations will happen beneath it. The system must decide which data should enter the context, which information must not be disclosed to a model, what remains current, what came from an unverified source, and what may be committed to memory.

In marketing, incorrect context can be more dangerous than a mediocre model. An excellent generator using an outdated price list will create a convincing but false advertisement. A strong AI copywriter that does not know a prohibited medical claim will increase regulatory exposure. An agent connected to the wrong data can automate a mistake faster than a human can notice it.

The end of the prompt era therefore marks the beginning of the context-infrastructure era.

Autonomy Without Control Is Technical Debt

In marketing, it is easy to be impressed by a system that "does everything by itself." It is harder to calculate the cost of an incorrect publication, data leak, noncompliant claim, uncontrolled budget, or campaign that scales the wrong signal.

OWASP identifies agentic-system risks including prompt injection, tool abuse, privilege escalation, data exfiltration, memory poisoning, goal hijacking, and excessive autonomy. Access to documents, websites, emails, and external systems expands the attack surface because malicious instructions can reach an agent indirectly as well as through direct user input.

A campaign agent might read a brief containing a hidden instruction, save a corrupted rule to memory, use a tool with overly broad permissions, or transmit confidential assets to the wrong service. Security policies therefore cannot be added as an afterthought inside a prompt. They must operate at the level of authorization, tools, memory, networking, secret storage, and logs.

The most important principle is simple: autonomy should increase with demonstrated reliability, not with the ambition of the sales presentation.

A practical deployment model looks like this:

• The agent prepares a proposal; a human approves the direction.
• The agent produces variants in approved formats; automated validators block violations.
• The agent can publish only to permitted channels and within defined limits.
• Any change to budget, audience, central claim, or use of a person's likeness requires approval.
• Every external action leaves a trace and can be tied to a specific policy version.
• After a sequence of failures, the process stops instead of retrying indefinitely.

OpenAI recommends retaining approvals for MCP tools, filtering inputs, applying guardrails, evaluating traces, and passing only structured fields between workflow nodes to reduce the propagation of malicious instructions.

These are not optional extras for cautious enterprises. They are the foundation of a product expected to operate without constant human supervision.

The Brand Will Become Executable Code

Brand guidelines have traditionally been documents written for people. An agent needs something more: rules that can be executed, validated, and versioned.

A tone described as "premium but approachable" may be meaningful to an experienced copywriter, but it is too ambiguous for automated publication. The system needs examples, permitted and prohibited terms, a messaging hierarchy, approved claims, logo rules, minimum contrast requirements, policies governing the representation of people, industry restrictions, and procedures for resolving conflicts.

Google is expanding text guidelines for AI Max specifically so advertisers can define prohibited terms and concepts in natural language. Salesforce describes goals, budgets, guardrails, and autonomy limits. Adobe is working to make an evolving brand context available to agents.

The shared direction is clear: brand governance is no longer limited to checking content after it has been generated. It is becoming part of the execution environment.

This opens a new category of technology. Organizations will need brand-policy repositories, regression tests for communications, automated validation across every modality, rule versioning, and mechanisms capable of proving which policy was in force when an asset was generated and published.

The strongest brand will not have one "magic prompt." It will have the best-designed decision system.

Provenance Is No Longer Optional

If a campaign produces hundreds of variants using multiple models, the organization must know where every asset came from and what happened to it along the way.

C2PA is developing Content Credentials: cryptographically secured metadata that links a file with information about its origin and modifications. The C2PA implementation guide covers elements such as digital-source classification, declarations of AI use, identification of regions modified by AI, and the ability to record data about the model, parameters, and ingredients used in generation.

In the European Union, Article 50 of the AI Act introduces transparency obligations concerning, among other matters, machine-readable marking of synthetic content and disclosure requirements for deepfakes and certain text published on matters of public interest. The European Commission states that the principal transparency obligations have applied since August 2, 2026, with transitional arrangements for some systems introduced earlier.

Not every AI-generated advertising image is a deepfake, and the obligations of a model provider are not identical to those of a company publishing content. The regulation should therefore not be reduced to the slogan "every AI advertisement needs a large label." Instead, the process should distinguish the type of content, the role of the entity, the intended use, the required disclosure, and editorial responsibility.

A campaign agent should preserve provenance automatically. If transparency depends on an operator remembering to add information manually at the end of the workflow, the system is not ready to operate at scale.

What Will Actually Happen to Creative Work

The simplest narrative says that the agent will replace the agency. It is a dramatic claim, but an incomplete one.

Agents will first attack the cost of coordination: rewriting briefs, manual versioning, formatting, localization, collecting files, repetitive amendments, reporting, and transferring information between tools. They will then take over some low-risk decisions for which clear feedback signals exist. Decisions with ambiguous objectives, difficult-to-measure success, and high reputational costs will be the slowest to automate.

The human role will shift from producing every variant to designing the system that creates and evaluates those variants. The strategist will define the problem and priorities. The creative director will define the visual language and aesthetic boundaries. Legal will define policies and risk thresholds. The AI operator will manage tools, context, and evaluations. The analyst will assess the quality of the signal the agent is optimizing against.

This does not mean that the idea will become less important. Quite the opposite. If production becomes inexpensive, average creative will be flooded by thousands of equally average variants. Competitive advantage will not come from the ability to generate another banner but from choosing the right tension, story, point of view, and moment.

AI will reduce the cost of execution. It will not guarantee meaning.

FOTOhub's Role in This Shift

For readers arriving through Crunchbase, the context should be stated precisely. FOTOhub is a platform developed in Poland for generating and managing multimedia content across images, video, audio, text, and design.

According to its own product materials, FOTOhub.app is evolving into a multi-model Creative AI OS. Its documentation describes Gabriel AI as a layer that classifies intent, selects a model, optimizes the prompt, estimates cost, and can return a multi-stage workflow. The platform's official announcement also describes workflows with durable execution, schedules, webhooks, human approvals, and node-level budgets.

This information comes from the company's own documentation and communications, not from an independent audit of performance. It should therefore not be used to claim that FOTOhub can already plan, publish, and optimize complete marketing campaigns autonomously. The available materials do, however, confirm components required to move in that direction: model routing, multimodal creation, an API, and multi-stage workflows with approval gates.

That distinction matters. The worst strategy for an AI startup is to sell a vision as if it were already an operational fact. The best strategy is to show which layers work today, which remain under development, how they are measured, and where human involvement is still required.

FOTOhub should not attempt to build "one supermodel that creates an entire campaign." A more durable position exists one layer higher: a system that understands the objective, preserves context, selects engines, connects modalities, controls costs, runs the process, and requires approval before taking external action.

In that architecture, Gabriel is not merely an automatic prompt writer. It can become the decision interface between user intent and a complete creative workflow. The image may be generated by one model, the video by another, the voice by a third, and validation and formatting by additional services. The user should not need to care which engine is fashionable. The user should care whether the system achieved the objective within the budget and in accordance with the law and the brand.

That is FOTOhub's opportunity. Not the number of models in itself. Not a catalog of provider logos. Not the promise that "AI will do everything."

The opportunity is to take responsibility for the path from objective to approved asset.

How a FOTOhub Agent Should Work

A creative-campaign agent could ultimately operate through a clearly constrained cycle:

1. Accept the objective. The user defines the product, market, audience, channels, budget, deadline, and metrics.
2. Build the brief. The system identifies missing information, asks questions, and does not proceed if critical constraints remain unknown.
3. Load brand context. It retrieves approved assets, tone, legal rules, examples, and prohibitions.
4. Map the plan. It breaks the campaign into formats, variants, languages, and dependencies.
5. Select the engines. It chooses models according to quality, cost, speed, data policy, and intended use.
6. Produce and evaluate. It generates versions, runs automated tests, and rejects outputs that fail to meet required thresholds.
7. Request approval. A human approves the strategy, central idea, product claims, use of likeness, and publication.
8. Deliver. Approved files are prepared for each channel or transferred through permitted integrations.
9. Preserve the trace. Every asset retains a history of the model, inputs, cost, version, validations, and approvals.
10. Learn from decisions. Rejections and revisions improve future recommendations but are not written unconditionally into memory as "the truth about the brand."

The most important part of this architecture sits between steps seven and eight. Production can be extensively automated. Publication must remain permissioned.

If an agent can generate one thousand variants but cannot explain its model selection, cost, or why it rejected 997 of them, it has not removed chaos. It has automated chaos.

The Greatest Illusion: More Content Means Better Marketing

Agentic production can radically increase the supply of marketing assets. It does not follow that those assets will perform better.

A system optimized for clicks may produce increasingly aggressive messaging. A system optimized for short-term acquisition cost may damage long-term positioning. An agent learning only from advertising-platform metrics may amplify attribution errors. Producing one hundred variants will not help if every one of them comes from the wrong brief.

A campaign agent therefore needs multiple layers of objectives:

• A channel outcome, such as cost or conversion.
• A business outcome, such as margin, retention, or customer quality.
• Brand constraints that cannot be sacrificed for a short-term improvement in a metric.
• Legal, ethical, and operational constraints.
• An exploration budget to prevent the system from endlessly cloning one temporarily successful motif.

McKinsey describes agentic AI in marketing as systems that plan, decide, and execute tasks across workflows with limited human input, while also arguing that the largest near-term gains are likely to come from redesigning processes around human - AI collaboration rather than pursuing automation without judgment.

That is the most credible perspective. An agent will not repair a bad strategy. It will only execute it faster.

When an Agent Should Stop

The most mature agent is not the one that never asks questions. It is the one that knows when further autonomy is no longer rational.

It should stop when:

• The brief contains conflicting objectives.
• Rights to use an asset or a person's likeness are missing.
• Costs exceed a defined threshold.
• A tool or model does not comply with the data policy.
• Validators cannot reach the required confidence level.
• The campaign operates in a high-risk domain.
• A goal-hijacking or prompt-injection attempt is detected.
• An action is public, costly, or difficult to reverse.
• Further retries fail to improve the result.

OpenAI recommends escalation after failure limits are exceeded and before high-risk actions. Anthropic recommends simple architectures, process transparency, and rigorous tool testing. OWASP highlights the dangers of excessive autonomy and tool misuse.

Autonomy without the ability to stop is not intelligence. It is a lack of control.

Who Will Win the Agentic Era of Marketing

It will not be the company with the longest prompt.

It will not necessarily be the lab with the strongest individual model.

Nor will it be the platform with the greatest number of integrations if users still have to select every engine, rewrite the context, and repair every handoff themselves.

The winner will be the system that combines five things most effectively:

• Context: Knowledge of the brand, product, audience, and decision history.
• Orchestration: Selection of models, tools, sequences, and alternative execution paths.
• Governance: Permissions, approvals, policies, and auditability.
• Evaluation: The ability to assess not only file quality but the entire chain of decisions.
• Distribution: Presence where the campaign is actually planned, produced, and launched.

OpenAI is commercializing the harness. Adobe is building an agentic content supply chain. Salesforce is building an objective and customer-data layer. Google is connecting agents to its own advertising channels. WPP is encoding agency knowledge and processes. Each company is attempting to control a different segment of the journey from intent to outcome.

For FOTOhub, the natural battleground is the creative execution layer: multimodal production, routing, workflows, cost control, and context preservation across models.

Technology alone, however, will not be enough. The agent must earn trust. Trust does not come from promising autonomy. It comes from predictability, explainability, control, and proof that the system knows when it should not act.

The Next Interface Will Not Be a Text Box

The prompt box became the icon of the first phase of generative AI. An empty field suggested infinite possibility, but it placed the full burden of defining the task on the user.

The next interface will look more like a command center.

The user will define the objective, budget, deadline, risk, and degree of autonomy. They will see the plan, sources, costs, progress, decision points, and reasoning behind critical choices. They will not need to know the syntax of every model. They will manage the operating policy of the system.

The prompt will remain one of the inputs. But it will no longer be the product, the workflow, or the competitive advantage.

In 2023, generating an image from text was an advantage. In 2024 and 2025, the market raced to improve model quality and add new modalities. In 2026, the essential question is whether AI can maintain an objective, use the correct tools, work across multiple stages, and execute safely in a real operating environment.

The end of the prompt era does not mean we will stop talking to AI.

It means we will stop explaining every action separately.

Instead of writing one hundred instructions for one hundred assets, we will define a single responsibility: build the campaign, follow these rules, stay within this budget, stop before publication, and show why every important decision was made.

When a system can do that, it is no longer a generator.

It becomes a participant in the process.

And the platform capable of giving that participant context, tools, memory, boundaries, and a complete creative workflow will not be another prompt-writing application.

It will be the operating system for agentic production.