Your Marketing Workflow Has Five Tabs Open. That Is the Problem.
If you are a marketer in 2026, the picture probably looks like this. ChatGPT for drafting. A separate tool for image generation. Another tab for keyword research. A scheduling app that does not talk to any of them. And in the middle, you, switching between tabs at least sixty times a day to push a single campaign live.
That is not an AI workflow. That is five disconnected tools pretending to be one.
What the most productive solo marketers in Hong Kong are doing in 2026 is different. They are building a small, deliberate stack of AI agents, four is the most common number, where each agent owns one phase of the workflow and hands off cleanly to the next. The stack runs $50 to $300 per month. It replaces work that used to require a small team.
This article walks through the exact four-agent structure, names the tools that actually work for each role today, and shows where it breaks if you skip the planning step.
Why the 4-Agent Marketing Stack Beats a Big AI Tool Stack
The trap most marketers fall into is collecting tools, not designing a workflow. The four-agent stack solves the opposite problem. You pick four specific roles, each one owns a clear stage of the campaign, and the output of one becomes the input of the next.
The four agents are: a Research Agent, a Drafting Agent, a Visual Agent, and a Publishing Agent. Together they cover the full pipeline from "we should write about X" to "the post is live and being measured". Nothing else lives inside this stack. Everything else, analytics, CRM, paid ads, is downstream.
The reason the four-agent number sticks is that adding a fifth agent almost always adds a hand-off problem rather than capability. Each hand-off between agents is where errors creep in. Four agents have three hand-offs. Five agents have four. The math of reliability gets worse fast.
Agent 1: The Research Agent
This agent's only job is to turn a fuzzy topic into a concrete brief. You give it a phrase like "we want to write about AI for HR teams this month". It comes back with the current angles being discussed, the search terms real people are typing, and a one-page brief that the Drafting Agent can use.
The tools that work for this role in 2026 are Perplexity Pro, Claude with Projects, or ChatGPT with web browsing on. The choice matters less than the structure. What matters is that your Research Agent runs the same playbook every time so the brief format is consistent.
Here is a copy-paste-ready Research Agent prompt that produces a structured brief:
You are my Research Agent. For the topic below, produce a brief with five sections:
1. Current angles (3 most-discussed framings of this topic in the last 30 days, with source links)
2. Search terms (5 real Google queries someone would actually type, ranked by intent)
3. Audience pain points (3 specific frustrations this article should resolve)
4. Unique angle (1 sentence stating the take only we could write)
5. Key claims to verify (any statistic or quote that must be fact-checked before publishing)
Topic: [PASTE TOPIC HERE]
Run that prompt once at the start of every content piece. The output becomes the input for Agent 2.
Agent 2: The Drafting Agent
This is the agent that turns the brief into a first draft. Its only job is writing in your voice, following your structural rules, hitting your target length. It does not research. It does not generate images. It does not publish.
The tool choice for the Drafting Agent matters more than for any other role because voice is hard to fake. Claude tends to produce drafts that need less rewriting if you give it a system prompt with three to five examples of your past writing. ChatGPT works well if you build a custom GPT and load your style guide as a file. Either works. Mixing tools mid-campaign produces inconsistent voice.
The mistake most solo marketers make is asking the Drafting Agent to do everything in one shot. Better: split the draft into three passes. Pass 1, structure only, just headings and one-sentence summaries per section. Pass 2, fill in the body. Pass 3, edit for voice. Three passes through a smaller scope produces a better draft than one mega-prompt asking for everything at once.
Build a Drafting Agent prompt that takes the brief as its full input. Tell the agent exactly which sections to write, how long each one should be, and what tone to use. The narrower the instructions, the cleaner the output.
Agent 3: The Visual Agent
Every piece of content needs a visual. A thumbnail, a hero image, social variants, a sticker for X. The Visual Agent handles all of it from a single prompt anchored in the brief from Agent 1.
The tools that work in 2026 are Midjourney V8 for hero images that need to feel premium, Flux for fast iteration when you need 20 variants in 5 minutes, and Ideogram if you need text rendered on the image without artifacts. Pick one as your default, do not jump between tools per project, because each tool has a learning curve and you want to invest in one.
The pattern that produces consistent visuals across a series is to write a "base scene" once in the Research Agent stage, then use the same base scene for every image in the series. Only the text overlay changes. This is exactly how brand magazines maintain visual coherence across a 40-page issue. The base scene locks the mood; the variable text labels the specific image.
A working base-scene template looks like this:
Bold editorial photograph of [SUBJECT]. [LIGHTING DESCRIPTION]. Colour palette: [3 SPECIFIC COLOURS]. Composition: [LEFT/RIGHT/CENTRE EMPHASIS], with ample empty space in the [TOP/BOTTOM] third for title overlay. No logo, no text in the image itself. Style reference: high-end magazine, not stock photography.
Save that template once. Reuse it across the whole campaign. Visual coherence is the cheapest signal of professionalism you can buy.
Agent 4: The Publishing Agent
The Publishing Agent is the one most marketers skip, and it is the most important one. Its job is to take the final draft and the visuals and turn them into platform-specific variants ready to schedule.
What that means in practice: a single article becomes a Facebook post, a LinkedIn variant in a different voice register, three X posts, a Threads version, an Instagram caption, and an email subject line plus preheader. The same source content, six different output shapes, each one optimised for the platform's algorithm and audience expectations.
The tool that has emerged as the de-facto Publishing Agent in 2026 is a combination of Make.com or n8n for the orchestration, plus a Claude or GPT-4o agent that handles the actual rewriting. Make and n8n are no-code workflow tools. You build the workflow once, drop in the source content, and the variants generate automatically.
The cost saving here is dramatic. A solo marketer publishing one article a week generates 6 platform variants. Across 50 weeks that is 300 variants. At an average of 15 minutes per manual variant, that is 75 hours per year of work the Publishing Agent removes.
Where the 4-Agent Stack Breaks
The stack works beautifully when set up correctly and fails reliably in three places when it is not. Knowing the failure points lets you avoid them.
Failure 1: Skipping Agent 1. Marketers love jumping straight to the Drafting Agent because writing feels productive. But a draft built on a fuzzy brief produces a vague article. The single highest-leverage hour in the whole stack is the one spent on the brief. Do not skip it.
Failure 2: Mixing voices between agents. If your Drafting Agent uses Claude and your Publishing Agent uses GPT-4o, the platform variants often drift away from the original voice. The fix is simple: lock the voice once at the Drafting Agent stage, then pass exact text snippets, not free-form rewrites, to the Publishing Agent. Tell the Publishing Agent to "preserve the original phrasing for the hook sentence and the conclusion".
Failure 3: Treating the visual as an afterthought. If you only think about visuals after the draft is done, you end up with stock-style images that do not match the article's argument. The fix is to brief the Visual Agent from the same source brief as the Drafting Agent. Both agents work from the same one-page reference. That is what produces text-and-image cohesion.
Try Building Your Own Stack This Week
Here is the smallest version of the stack you can test in three days. Pick one upcoming article. Spend Monday writing the brief with Agent 1. Tuesday, draft with Agent 2 using a three-pass approach. Wednesday, generate visuals with Agent 3 using a single base scene. Thursday, run the Publishing Agent flow to generate the six platform variants. Friday, schedule them all and watch what publishes.
Total time: about four hours spread across the week. The same article via the old, all-manual approach would take eight to ten hours and produce less consistent output. The four-agent stack is not magic. It is just deliberate sequencing of tools you probably already have access to.
The bigger shift is the mental one. You stop thinking of AI as "a tool I open to do a task" and start thinking of it as "a team of narrow specialists each doing one thing well". That is how solo marketers operate at the output level of a five-person team in 2026. We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth.
Ready to Build Your AI Marketing Stack?
You now have the blueprint. The hard part is wiring the four agents together so the hand-offs are clean and the voice stays consistent. We'll walk you through every step, from selecting the right tool for each role to building the no-code orchestration that ties them together.
