Your content team:
9 agents and one human.

The uncomfortable bit

AI content isn’t slop. Loopless content is.

When a marketing team brings us their content problem, it is rarely a lack of output. They have ChatGPT, they have a calendar, and they post. What they don’t have is anything that remembers. Generate, post, forget, and next week starts from a blank page. The 50th post is as blind as the first.

The install we run replaces that line with a loop: research, make, approve, post, flag, boost, measure, repeat. Nine stages, one human checkpoint, one budget cap. The rest of this page walks the loop in the order we build it.

The full system

One loop, 9 stages

We build the nine stages in order, because each one feeds the next. Brand and research come first so the agents have something true to work from. Production, approval, scheduling and publishing turn that into posts. The last three stages, flagging winners, boosting them and collecting the data, are what make the second month better than the first.

Across all nine, the agents do the grunt work. Your team keeps two jobs: approving what goes out and capping what gets spent. For the company, that is a content operation that runs every week without a new hire, and gets more accurate the longer it runs.

Stage 01 · Brand

Feed it a brand brain

Agents write like nobody until you tell them who you are, so the first thing we build is the brand doc. It has three parts. Positioning covers mission and vision, the enemy, the business mantra, niche and USP. Branding covers guidelines, moodboard, tone of voice and recordings. The ICP names the primary buyer and two avatars.

Founders rarely have this written down, so we start with a voice note. Record one on a walk, talking without stopping about how you see the business, and the agent turns it into positioning, mission, USP and avatar drafts. You sign off the positioning. We still pay talented people for the guidelines and the moodboard, because that is where the brand gets judged.

From here on, every agent in the loop reads this doc before it writes a word. So can every new hire.

Stage 01 · Audit and competitor research

Build off real people, not AI

Before anyone writes, the research agent scrapes what is already winning in your niche. Apify’s MCP server pulls top posts from Instagram, TikTok, X and thousands of other sources. Reddit and Google Trends add the real questions and complaints, in your buyers’ own words.

The agent clusters the hooks that keep winning, from your account and your competitors’. You decide which channels make the cut. Four of eight ticked is the usual shape of that decision: produce for the channels your buyers are on and leave the rest alone.

Stage 02 · Content strategy

Pick the formats first

Research tells you what to say. Strategy decides where it lives. We map eight content types, from blog posts and email through to webinars and video, and the agent assigns each research insight to a type and drafts the monthly plan.

You decide which two or three types the team can actually sustain. That limit matters more than the plan. A system that promises eight types and ships two loses the team’s trust. One that commits to two and ships them every week earns it.

Stage 02 · Instagram format playbook

What goes in each format

Every Instagram format has jobs it is good at, and an agent without a map will post memes as Stories. So the strategy agent carries a cheat sheet: single posts for checklists, quotes and promotions; Reels for tutorials, before-and-afters and reviews; carousels for case studies, how-to guides and blueprints; Stories for polls, countdowns and launches.

The agent tags every idea with a format before anything is made. You set the weekly mix, how many Reels against how many carousels, and change it when the data says to.

Stage 03 · Content production

The factory floor

Production is where the volume comes from, and where we are strictest about which model does what. Higgsfield makes the video and visuals for Reels and ads. Nano Banana Pro, Google’s image model, runs through Gemini. GPT-Image-2 gives a second look on stills. Claude Opus 5.5 writes the scripts, captions, blogs and emails in your voice, with the brand doc open.

The agent writes the brief, picks the model, generates variants and files them for review. At this stage you decide nothing. Nothing from the factory floor reaches a channel until a person has seen it, which is the next stage.

This is what comes off the floor: video ads and product stills from the production stage, shown here as silent preview loops.

Stage 04 · Approval and review

A human in the loop

Nothing goes live without a yes. We put the review queue in whichever board the team already lives in, Notion, Google Sheets or ClickUp, so approving a post is a column change in a tool the team opens every day anyway.

The agent files each draft with its asset, caption, channel and the research that justified it, then applies your edits. You approve, edit or kill. Edit notes feed back into the brand doc, so a correction made once holds for every post after it. For most teams this is the one daily touchpoint with the system.

Stages 05 and 06 · Schedule, then publish

Approved? It ships itself

Approval flips the status. The agent pushes the post to the scheduler, and the scheduler publishes everywhere the plan says: Buffer, Postiz or Sprout Social, out to Facebook, Instagram, X, TikTok, YouTube, Telegram, WhatsApp and LinkedIn.

The agent picks slot times and formats per channel and schedules through the tool’s API. You decide which channels are in play this quarter. Postiz is open source, so a team that wants the posting layer on its own servers can have it there.

Stage 07 · The winners get flagged

Let the data pick favourites

Once posts are live, the loop starts paying for itself. Every post is scored against the account’s own average, and the outliers get a flag. No opinions, no office politics.

The agent pulls views, saves, shares and comments daily and flags anything well above normal. You decide what counts as a winner and what a winner is worth. Inside the company this replaces the Monday argument about which post did well with a number everyone can see.

Stage 08 · Boost the organic winners

Only pay for what already won

Organic is your free A/B test. A flagged post has already proven its hook with people who chose to watch it, so it becomes an ad. The agent rebuilds the post as a Facebook and Instagram ad, launches it on a capped test budget and reports cost per result.

You set the cap. Always. Spend is the one thing an agent never sets in our installs, and that rule is what makes the loop safe to leave running.

Winner → adFacebook + Instagram, small test budgetLive

1. flagged post
2. ad built
3. capped test
4. report

Agent does

Rebuilds the post as an ad, launches it on a capped test budget, reports cost per result.

You decide

The budget cap. Always. Spend is the one thing an agent never sets.

Stage 09 · Collect as much data as possible

Hoard the numbers

One analytics dashboard for organic and paid together. The agent pulls every platform and ad account into one view, every day, and tracks hook and format, channel, post time, saves and shares, ad cost per result and leads.

You decide which number is the north star. The more the system sees, the smarter the next loop gets, because next month’s plan is drawn from this view rather than from memory.

The loop

Feed it back. Start again.

All that data flows back into content strategy. Next month’s plan starts with last month’s winners: the hooks that got flagged, the formats that earned their slot, the channels that produced leads.

The agents are the same models anyone can rent. What the company ends up owning is the loop and the data inside it.

What it means for your company

One person runs it. The loop does the rest.

Nine stages, one human checkpoint, one budget cap. In practice one person runs the weekly content operation: they approve in the board they already use and set the spend behind posts that have already won. Research, drafting, design, scheduling, publishing, scoring and reporting happen without them.

When we install it we go in order: brand doc and research first, then production and the approval board, then the scheduler, and the flag-and-boost loop once there is a month of posts to score. Every lap after that starts smarter than the one before.