A few months ago, a client asked us a question that sounds simple and is not: "Can you just handle our SEO and content, properly, without us having to chase anyone for updates?"
The honest old answer was a retainer — a person or a small agency team working a handful of hours a week, reporting monthly, moving at the pace of whoever had time that week. We decided to try something different: instead of hiring a person to do the job, we built a team of AI agents to do it, and gave the client a window straight into the work.
This is what we built, why it is structured the way it is, and what running it has actually taught us.
Not One Agent — Thirteen, Each With a Job
The instinct when people hear "AI agent" is to picture one assistant that does everything. That is not how we built this, and it is not how any system that has to survive contact with a real business should be built. Instead, the client's marketing operation runs on thirteen specialised agents, each with a narrow, well-defined responsibility:
- Keyword Intelligence — mines Search Console and seed data to find and score opportunities
- Content Strategist — turns scored opportunities into content briefs
- Content Producer and Content Editor — draft and then score and refine pages before anything goes live
- Technical SEO — audits the site for issues and queues fixes
- SERP Intelligence — tracks how the pages actually rank and how competitors move
- AI Search / GEO — checks how the brand's entities and facts show up in AI-generated answers, not just blue links
- Authority / Digital PR — finds and queues outreach prospects for backlinks and mentions
- Local SEO, Conversion SEO and Analytics — watch demand, funnel performance and the numbers that actually matter commercially
- Learning — closes the loop by attributing results back to what the other twelve actually did
- SEO Commander — sits above all of them, sequencing work and deciding what runs next
Each one is good at exactly one thing. That is deliberate. A single generalist agent asked to research, write, audit and report tends to do all four passably. Thirteen specialists, each with a tight scope, do their one thing well — and more importantly, each one's output becomes a clean input for the next.
The Part That Actually Matters: They Hand Off to Each Other
The interesting engineering is not any single agent — it is the chain between them. Keyword Intelligence scores an opportunity. Content Strategist turns it into a brief without anyone copying a spreadsheet into a Slack message. Content Producer drafts against that brief. Content Editor scores the draft and either queues a fix or waves it through. Technical SEO separately audits the live site and feeds its findings back into what gets prioritised next.
None of that requires a human to shuttle information between four different tools and three different people, which — if we are honest — is where most marketing operations actually lose their time. Not in doing the work, but in relaying the work.
Autonomous Where It's Safe, Gated Where It Isn't
This is the question every client asks eventually, and it should be the first question any business asks before deploying something like this: what is the agent allowed to do without asking?
Routine, reversible, low-stakes actions run on their own — scoring a keyword, drafting a brief, flagging a technical issue, syncing a ranking number. Anything that publishes content, spends budget, reaches out to a third party for a backlink, or changes something live on the site sits in a queue waiting for a person to look at it and click approve. On the dashboard we built for this client, that queue is visible at a glance — right now it shows four items awaiting approval and nine open actions across the system, next to a button that says "Run all agents" that nobody gets to press without knowing exactly what it will trigger.
That distinction — autonomous for the reversible, gated for the consequential — is the actual engineering work. Anyone can wire an AI model up to a CRM and a publish button. Deciding, system by system, which of those two buckets each action belongs in is what separates something a business can actually trust with real operations from something that looks impressive in a demo and gets switched off after the first mistake.
Visibility, Not a Black Box
The client does not get a monthly PDF. They get a live monitor — every agent's status, its last run, what it produced, refreshed every second — so they can see at 9am on a Tuesday exactly what happened overnight: which keywords got scored, which pages got audited, what is sitting in the approval queue waiting for their decision. If an agent has been idle for three days because it genuinely had nothing useful to do, that shows too. We would rather show an honest "idle" than fabricate busywork to look active.
That transparency does something else worth mentioning: it builds trust in the system faster than any explanation of the architecture could. Watching the dashboard tick over in real time, and seeing a human approval request land with full context attached, tells a business owner more about whether to trust an agent than a slide deck ever will.
What We Have Learned Running It
The handoffs are harder than the agents. Getting one agent to research a topic well is the easy part now. Getting its output into a shape the next agent can act on without a human translating in between is where the actual engineering time goes.
Not every agent earns the same cadence. Some genuinely have daily work. Others — Local SEO and Analytics, in this client's case — are honestly waiting on more data before they have anything meaningful to act on, and the system should say so rather than manufacture output to look busy.
The approval queue is the trust mechanism, not a bottleneck. We initially worried a human-in-the-loop gate would slow things down. In practice it is the opposite: because the client knows nothing consequential happens without their sign-off, they are comfortable letting far more of the routine work run fully autonomous.
What This Means If You Are Considering Something Similar
You do not need thirteen agents on day one. This client's system grew from three to thirteen over several months, each one added once the last proved itself and a genuine gap showed up in what the humans were still doing manually. Start with the one process costing you the most hours, build it properly — scoped, observable, gated where it needs to be — and let the system earn its way to the next agent rather than designing the whole org chart up front.
Harzotech designs and builds multi-agent AI systems like this one for businesses across Nigeria and Africa — not a single chatbot, but a coordinated team with real visibility and real guardrails. Book an AI Agent Audit and we will map what a system like this would look like in your operation.