Building in Public: How Our AI Agents Run a Content Factory

This week in numbers (real, from our system)

• 🤖 AI agents running: 19
• 📝 Content published: 65 (blog RU 22, EN 23, Altezza 20)
• ⚙️ Generated programmatically: 4
• 📥 Leads in the system: 261 (+0 in the last 7 days)

Figures as of 2026-08-19 — computed by code from the DB and files, no manual entry.

Short answer: We run a content factory powered by AI agents that handle research, drafting, SEO optimisation, internal linking, and publishing — with a small human team steering strategy and quality. The result is a sustainable, scalable content operation where people focus on thinking, not on repetitive execution. Approximate throughput depends on scope and integrations, but the model is replicable for most small and mid-sized businesses.

Most founders talk about AI in the abstract. I'd rather show you the machine while it's running.

This is a living diary — updated weekly — of how the Arxitek team built and operates a content factory driven by AI agents. No polished case study. No invented numbers. Just the real workflow, the honest friction, and the decisions we make along the way. If you run a small or mid-sized business and you're wondering whether content automation is worth the investment, this post is written for you.

What Does "Building in Public" Actually Mean for a Tech Team?

Building in public is a practice borrowed from indie developers and startup founders: you document your process openly, share what's working and what isn't, and let your audience watch the product take shape in real time. For a software or AI team, it means resisting the temptation to only publish polished success stories and instead showing the scaffolding — the failed prompts, the pipeline rewrites, the agent behaviours that surprised us.

Why does this matter for business readers? Because transparency is the fastest shortcut to trust. When you see the actual workflow — not a sanitised demo — you can judge for yourself whether the approach fits your context. You can steal the good ideas and skip the dead ends we already explored.

For Arxitek specifically, building in public serves a second purpose: it forces us to document what we're doing in a way that's legible to non-engineers. Our clients are business owners and marketing directors, not AI researchers. If we can't explain our content automation stack in plain language, we probably don't understand it well enough ourselves. Writing this diary is, in a real sense, a quality-control mechanism.

The practice also keeps us honest about the limits of the technology. AI agents are powerful, but they are not magic. They make mistakes. They need guardrails. They require human judgment at the edges. Every entry in this diary reflects that reality — and if an entry ever sounds too smooth, that's a sign we're not looking hard enough at what went wrong.

Finally, building in public creates a feedback loop with the people reading along. Founders and marketing directors who follow this series regularly surface questions and edge cases we hadn't considered. That input has already shaped how we've tuned several agent behaviours. So this isn't just documentation — it's a collaborative process, and you're part of it.

How Do Our AI Agents Actually Work?

Let me walk you through the architecture without drowning you in jargon.

The content factory is a pipeline of specialised AI agents, each responsible for a discrete task. Think of it less like a single robot doing everything and more like a well-structured team where each person has a clear lane.

The Research Agent

Every piece of content starts with a research brief. The research agent takes a topic or keyword cluster, queries multiple sources, extracts relevant facts and angles, and returns a structured brief: key questions the article should answer, competing content to be aware of, and a suggested structure. This is the stage where we invest the most in prompt engineering — garbage in, garbage out applies here more than anywhere else.

The agent does not invent statistics. It surfaces real sources and flags anything it cannot verify. We have a hard rule: if a figure can't be traced to a primary source, it doesn't go into the article. That rule is baked into the system prompt, not just the style guide.

The Drafting Agent

Once the brief is approved — and a human does approve it — the drafting agent writes the article. It follows a tone-of-voice document, a banned-phrases list, and a structural template appropriate to the content type (explainer, diary entry, comparison guide, landing page). The draft is written in Markdown and deposited into a shared workspace for review.

The drafting agent is not trying to replace a writer. It is trying to eliminate the blank-page problem and the mechanical first-draft labour so that a human editor can focus on judgment: is this angle right? Is this section too long? Does this conclusion actually land?

The SEO and Internal Linking Agent

After the draft passes editorial review, an SEO agent runs a pass: it checks keyword density, suggests adjustments to headings, identifies internal linking opportunities within the existing content library, and generates a meta description and slug. It also flags any sections where the content is thin relative to the target keyword's search intent.

This is where content automation pays for itself most visibly. Manual SEO audits of every article are time-consuming and easy to skip under deadline pressure. An agent that does it automatically on every single piece means the baseline quality of our SEO never slips because someone was in a hurry.

The Publishing and Analytics Agent

The final agent handles publishing to the CMS, schedules social distribution, and sets up tracking. Post-publication, it monitors performance signals — ranking movement, engagement, scroll depth — and surfaces anomalies for human review. If an article drops in ranking, the agent flags it and suggests a refresh brief. The human decides whether to act on it.

What Does the Human Team Actually Do?

This is the question I get most often, and it's the right question to ask.

Here's the thing: the human team does everything that requires genuine judgment, taste, and accountability. We set the editorial strategy. We decide which topics to pursue and which to ignore. We approve every brief before it goes to the drafting agent. We edit every draft before it publishes. We make the call on whether a piece of content is honest, useful, and consistent with our values.

What we don't do anymore is the mechanical work: the first-draft typing, the keyword spreadsheet maintenance, the manual internal link audits, the scheduling, the basic performance monitoring. AI agents handle that layer. And freeing ourselves from that layer is what makes it possible to run a content operation at scale without burning out or hiring a large team.

Look — I want to be direct about something. There is a version of this story where the founder says "we replaced our content team with AI" and frames that as a win. I don't think that's the right frame, and it's not what we did. The people on our team do more interesting, higher-leverage work now. The AI agents handle the routine execution so that the humans can focus on the thinking. That's the model I believe in, and it's the model that actually produces better content over time.

The Weekly Rhythm

Every Monday, the editorial lead reviews the content calendar and approves the week's research briefs. By Tuesday, the research agent has returned structured briefs for each piece. Wednesday is editorial review and brief sign-off. Thursday, drafts are in the workspace. Friday is editing, SEO pass, and scheduling. The cycle is tight, repeatable, and — critically — it doesn't depend on any single person being available every day.

What Has Broken Along the Way?

Building in public means telling you about the failures, not just the wins.

Prompt drift. Over time, without active maintenance, agent outputs drift from the original intent. A drafting agent that was producing crisp, punchy copy gradually started producing longer, more hedged sentences because we had approved a few longer pieces in a row and the feedback loop reinforced the wrong behaviour. We now do a monthly prompt audit — a structured review of recent outputs against the original tone-of-voice document.

Hallucinated sources. Early in the build, the research agent occasionally returned plausible-sounding but unverifiable citations. We caught this during editorial review, but it was a reminder that the "no invented data" rule needs to be enforced at the system level, not just the style guide level. We rewrote the research agent's prompt to require source URLs for every factual claim and to explicitly flag anything it cannot verify.

Over-automation of judgment calls. We experimented briefly with giving the SEO agent the authority to auto-publish after a certain confidence threshold. It was a mistake. Publishing decisions involve context that an agent doesn't have access to — a news event that makes a topic sensitive, a client relationship that makes a particular angle awkward, a strategic pivot that changes the content priorities. Publishing authority stays with humans. Full stop.

Integration friction. Connecting agents to the CMS, the analytics platform, and the social scheduler introduced more edge cases than we expected. Files in the wrong format, API rate limits, timezone mismatches in scheduling. None of these were insurmountable, but each one cost time to diagnose and fix. If you're planning a similar build, budget generously for integration work — it's where the hidden complexity lives.

Is Content Automation Right for Your Business?

Let's be real: not every business needs a multi-agent content pipeline. The infrastructure has a setup cost — in time, in engineering attention, and in the ongoing maintenance of prompts and integrations. If you're publishing once a month and you have a capable writer, the overhead probably isn't worth it.

But if any of the following are true for your business, content automation deserves serious consideration:

• You need to maintain a consistent publishing cadence but your team's bandwidth is unpredictable.
• You have a large library of existing content that needs regular SEO refreshes.
• You're running content in multiple formats simultaneously — blog, social, email, landing pages — and the coordination overhead is killing you.
• You want to scale content output without scaling headcount proportionally.
• You're spending a disproportionate amount of your marketing team's time on mechanical tasks rather than strategy.

The honest answer is that content automation is a force multiplier, not a silver bullet. It amplifies the quality of your editorial strategy if that strategy is sound. It amplifies the gaps if it isn't. The technology is ready. The question is whether your process and your editorial judgment are ready to drive it.

A Practical Comparison: Manual vs. Agent-Assisted Content Operations

Dimension	Manual Content Operation	Agent-Assisted Content Operation
First-draft production	Writer starts from blank page	Agent produces structured draft from approved brief
SEO audit	Manual, often skipped under deadline	Automated on every piece, every time
Internal linking	Inconsistent, relies on writer's memory	Systematic, drawn from full content library
Publishing & scheduling	Manual steps, error-prone	Automated with human approval gate
Performance monitoring	Periodic, reactive	Continuous, anomaly-flagged for human review
Human focus	Split between execution and strategy	Concentrated on strategy, judgment, editing
Scalability	Headcount-constrained	Process-constrained (easier to fix)

Neither column is universally better. The right choice depends on your volume, your team structure, and your appetite for building and maintaining the pipeline.

Frequently Asked Questions

What is "building in public" and why does it matter for AI projects?

Building in public means documenting your process openly — including failures, pivots, and unresolved problems — rather than only sharing polished outcomes. For AI projects specifically, it matters because it builds trust with a sceptical audience, forces clearer internal documentation, and creates a feedback loop with the people who will eventually use or buy what you're building. It's a discipline as much as a communication strategy.

How do AI agents handle content quality control?

AI agents can enforce structural rules, style-guide compliance, keyword requirements, and source-verification checks automatically and consistently. What they cannot reliably do is judge whether a piece of content is genuinely useful, honest, or strategically appropriate. That judgment stays with the human editorial layer. The best content automation setups treat agents as a quality floor, not a quality ceiling — they prevent the obvious mistakes, freeing humans to focus on the harder calls.

What does content automation typically cost for a small business?

Costs vary significantly depending on the tools, the number of integrations, and whether you build custom agents or use off-the-shelf platforms. On average across the market, businesses working with specialist teams report investment ranges from a few hundred to several thousand dollars per month, depending on scope and complexity. The more useful question is cost per published piece relative to your current manual cost — that comparison almost always favours automation at any meaningful publishing volume.

Can AI agents replace a content team entirely?

In practice, no — and the businesses that try to run fully autonomous content operations without human oversight tend to produce content that is technically compliant but strategically hollow. AI systems help businesses scale execution; they don't replace the editorial judgment, brand voice ownership, and strategic thinking that a content team provides. The realistic model is a smaller human team doing higher-leverage work, supported by agents handling the mechanical layer.

How long does it take to set up an AI-powered content pipeline?

A basic pipeline — research brief, draft, SEO pass, publish — can be assembled in a matter of weeks if you're working with an experienced team and your CMS has reasonable API access. A more sophisticated operation with custom agents, tight integrations, and tuned tone-of-voice enforcement typically takes one to three months to reach a stable, production-ready state. Budget additional time for the first month of operation, when prompt tuning and edge-case fixes are most frequent.

Where We Go From Here

This diary is a living document. Each week, we add an entry covering what changed, what broke, what we learned, and what we're building next. The goal is not to present a finished system — it's to show the system as it actually evolves.

If you're a business owner or marketing director thinking about content automation, the most useful thing I can offer is not a polished pitch deck. It's this: an honest, ongoing account of what it actually takes to run AI agents in production, at the level of detail that lets you make a real decision.

If something in this post raised a question specific to your situation — your industry, your team structure, your existing tools — get in touch. We're happy to talk through whether this model makes sense for what you're building.