How I built a virtual agency that manages your content pipeline.
My Ubuntu laptop is running hot. A Solana Seeker phone — one of the first batch, bought because I believe in the chain — is plugged into it via USB. On the phone screen, I can see Instagram opening by itself. The cursor moves, taps, navigates to the post dialog on the MemberMagix account. Droidclaw and Sonnet are driving.
I’m filming it with my iPhone. This is it. The first autonomous post.
Then it gets lost. The agent fumbles somewhere in the Instagram UI, and I stop the recording. I turn back to my development environment to figure out what went wrong.
What I didn’t realize: Sonnet kept going.
These models, when you give them a task, they want to finish. I wasn’t watching the phone anymore. I was reading logs, checking configurations, trying to debug the flow. A few minutes pass. I glance to my right, and there it is — a selfie of me, posted to Instagram. No caption. No hashtags. Just my face, caught by the phone camera at some angle while I was hunched over the laptop, completely unaware.
I archived it immediately. I thought I’d lost the photo — it was taken with my iPhone off the Android screen, raw and unplanned. But it survived. And honestly, it tells the whole story better than anything I could have staged: the mess, the ambition, the gap between what you plan and what actually happens when you let an agent loose.
The Vision
The idea behind all of this is straightforward: automate the entire content pipeline. From client brief to campaign strategy, from individual posts to publishing on the platform. No human intervention. Fully autonomous.
Solutions exist, loosely. But none of them are good. And I gave myself a constraint that made this harder on purpose: no platform APIs. No Graph API, no Facebook SDK. I wanted this to work like a human would — opening the app, navigating the interface, posting. Partly because that’s the harder, more interesting problem. Partly because features like Instagram Reels music selection simply aren’t available through the API.
Charcoal International
I’ve always been a fan of Mad Men. That show about an advertising agency in 1960s Manhattan, Don Draper in his corner office, running campaigns that shaped culture. My vision was simple: build a Don Draper for 2030. The personality of a seasoned creative director, the tools of the near future.
I spun up the agent using my Agent Builder on the Ubuntu machine. He needed a name. I started with the color: one of the leading ad agencies worldwide is Grey. From Grey, I got to charcoal — which happens to be a color in the surfstyk identity. The agency became Charcoal International. The agent became Sharky — totally random, but it stuck.
And it works, because Sharky is a character that goes for it. He doesn’t hesitate, doesn’t overthink, doesn’t ask for permission twelve times before making a move. He’s got a profile picture, a personality brief, and a mandate to run campaigns.
The Structure
Behind the fun naming sits a serious architecture, built on years of running marketing projects and managing delivery teams.
The system works with clients — accounts, essentially. Each account has its own manager agent. Each client comes with branding guidelines, a creative brief, objectives, target platforms, tone of voice. The whole playbook that any real agency would build before creating a single piece of content.
The first project I put through the image pipeline was MemberMagix — my WordPress plugin for lightweight membership protection. I needed to test the automation end to end, and MemberMagix needed an Instagram presence. That’s where the selfie incident happened. A chaotic first test, but proof that the system could navigate a real phone, interact with a real app, and push content — even if it picked the wrong content.
Kong Quant and the Video Pipeline
With the image flow validated, I moved to the real challenge. Kong Quant — a market analysis project I’ve been running. A gorilla in a business suit and nerdy glasses who scans up to a thousand assets every four hours and tracks the market. He gives away a daily Prime on X: a condensed snapshot of a ticker’s chart, analysis, and indicator data.
Kong already had an automated X account — X actually allows this, as long as you’re transparent. His account is labeled “automated by surfstyk,” and I’ve never touched it beyond the bio. He posts what he wants to post. I don’t delete anything.
I wanted Kong on Instagram and TikTok. Specifically, I wanted short-form video. Reels, TikToks. By experience, producing even a basic reel takes one to four hours of manual work. Sometimes a full day if you want it right. Automating that production was the ultimate goal.
On paper, the ingredients are all there. Image generators, video generators, Heygen, Synthesia. For Kong, I wanted an animated version of the character presenting market data. A 15-to-30-second short-form video with a text block for the post caption.
Easy on paper. In execution, a lot of things you have to consider. Consistent character animation at any quality level is ambitious, and even if you’re willing to spend two to five euros per post, you’ll burn time testing before you get it right.
We landed on a pragmatic approach: a fixed animated video template as the foundation, with dynamic text overlays on top. The structure of Kong’s content is consistent — a ticker symbol, a chart snapshot, analysis data — which makes a template viable. The data and the chart image come from an API I’d already built. ElevenLabs handles the voiceover. A lot of testing, a lot of tweaking to get the timing, placement, and tone right.
The result: a production pipeline that creates a finished short-form video for 20 to 30 cents. All API costs included. Compare that to one to four hours of manual production. The economics aren’t even close.
During the image testing phase, I also tried Minimax as an alternative model. Noticeable quality difference compared to Sonnet, but the pricing is attractive and it gets the job done. Worth keeping an eye on as these models evolve. For production, I stayed with Sonnet.
Five Minutes of Glory
Kong Quant’s first real video post was ready. The video came through clean — on brand, not lip-synced, a little raw but authentic for a fully automated production. I picked the music directly on Instagram while posting, a deliberate design choice. Sharky and I decided early on: don’t bake music into the video. Platform-native music avoids licensing and copyright headaches entirely. And if you catch a trending sound, it helps reach. Let the platform handle what the platform is good at.
I posted it. Watched it. Felt genuinely good about it.
Five minutes later, Meta suspended the account.
I appealed. Gave them everything — phone number, email, context. Explained that this wasn’t a spam operation, that there was a real person behind it with transparent intentions, that the content delivered actual value. Two hours later, the verdict came back: permanent. Gone for good.
Instagram doesn’t appreciate AI-generated content, especially on new accounts with no history. The finance angle adds another flag — crypto-adjacent content triggers their moderation systems hard, and I understand why. The space is full of people with short-sighted intentions, scams dressed up as analysis. I think blockchain technology has a genuine future, especially in the context of AI agents transacting with each other. But Instagram doesn’t share that view, and for a fresh account with zero social proof, you don’t get the benefit of the doubt.
I was frustrated. Took it personal. But then I talked to Sharky, and he wasn’t worried at all. “That’s normal,” he said. “Instagram shuts down small accounts like this before they even get a chance to grow. It’s in their community guidelines.”
The Pivot
TikTok is friendlier territory. They appreciate automated and AI-driven content — it’s cutting edge, it’s tech-forward, it aligns with their platform identity. There’s even a community around financial content on TikTok: FinTok. The audience is there.
The content machine stays the same. I just added a platform flag to the automation. This matters more than you’d think — each platform has different safe areas on screen. You can’t place important information where a like button or a comment icon will cover it. The template adapts based on whether the output targets Instagram, TikTok, or another short-form platform.
I spun up the TikTok account, adapted the template, and my day ended with a post on TikTok. One or two minor issues — I was too tired to fix them. Nothing structural. The first post is out, and as of today, TikTok hasn’t complained.
The Real Deliverable
Here’s the thing I learned that matters more than any technical detail: the delivery endpoint changes everything.
My original vision was full autonomy — the agent posts, engages, handles everything on the phone. I had to realize that this is a project on its own. The operational complexity of controlling a phone, navigating platform UIs that change constantly, dealing with terms of service that explicitly prohibit automation — it’s a rabbit hole I might explore later, but it’s not where the value is right now.
The actual posting, once you have the video and the caption with tags, takes under a minute. Pick the music, paste the text, publish. That’s a task you can do yourself or hand to a team member. The hard part — the creative production — is what takes hours. That’s what the automation solves.
So I changed the delivery endpoint. Working with different Claude session — I think of them as teams, the front-end team, the back-end team, the automation team — we landed on Telegram. Easy to set up, developer-friendly, and Sharky already had his own bot from the system architecture. Each client, even each product, can have a dedicated Telegram receiver. From the client’s perspective, the agency does the work and delivers a message: visuals — one to ten images or a video — plus a text block formatted as a code block so you just tap to copy. No selecting, no formatting. Tap, paste, post.
We tweaked the delivery to avoid compression on visuals and to make the text element as frictionless as possible. The result is a deliverable that lets anyone post quality content in under a minute.
That’s the real product. Not full autonomy — not yet. A production pipeline that turns a brief into a ready-to-post package, a full series of campaigns, delivered to your phone, for cents instead of hours.
What This Changes
Working with these models taught me something that applies far beyond content automation. They’re exceptional at focused, well-scoped tasks. Image generation, voiceover, data processing, text composition — individually, each step is remarkably capable. But the orchestration, the context switching, knowing where to draw the line between automation and human judgment — that’s still the human’s job.
A fool with a tool is still a fool. You need to know what you’re building before you start building it. And sometimes you need to build it wrong first to understand where the actual boundaries are.
Charcoal International is operational. The structure handles onboarding new clients, producing on-brand content across formats, and delivering it ready to post. For small businesses — the kind I work with — this changes the economics of content completely. Professional, consistent, on-brand social media presence without the agency retainer or the hours of manual production.
The Instagram ban stung for about two hours. Then it became the best lesson of the sprint: the technology works. The platforms are still catching up to what that means. Some of them, like X and TikTok, are already there. The rest will follow or diminish.
And that random selfie — the one I thought I’d lost — turned out to be the most memorable moment of the whole project. Me, unaware, caught by my own agent, posted without my permission. That is scary, beautiful, totally fun at the same time.
