Engineering Journal ยท Roadmap
Our Marketing Team Is an Agent Swarm - and I'm on the Send Button
A studio with a dozen products has a dozen stories to tell and no time to tell them. So the next thing we're building isn't another product - it's the team that markets the others. A swarm of specialised AI agents that research, write, and propose outreach around the clock, with one hard rule: the swarm drafts, but a human approves every message that reaches a real person. This is where we're taking it, and why the boring guardrail is the whole point.
Why a swarm, not one big agent?
One agent asked to "do our marketing" does everything averagely and nothing accountably. A swarm is a small team of narrow specialists, each good at one job and each inspectable on its own. A Scout watches for prospects, trends, and moments worth engaging. A Writer turns those into drafts in our voice. An Engager proposes the reply, comment, or message for a specific channel. A Qualifier scores who is actually worth my time and routes the hot ones up.
The same reason we build software as roles and modules instead of one giant function applies here. Narrow agents are easier to trust, easier to correct, and easier to switch off when one of them starts producing nonsense. A marketing "brain in a box" is a black box; a swarm is a set of jobs you can watch.
The rule that makes it legitimate: the swarm drafts, it does not send
Here is the line, and it isn't a nice-to-have. Every agent produces proposals, not actions. Nothing that touches a real human - a reply, a DM, an email - goes out on its own. Each proposed send stops at a gate and fires a webhook to me: on my phone I see the draft, the target, and the context, and I approve, edit, or reject. Only then does it leave. The agents run 24/7; the send button is still a person.
That single design decision is the difference between a growth engine and a spam cannon. An automated system that messages strangers unattended, at machine speed, pretending to be a person, is exactly the thing that gets accounts banned, laws broken, and trust burned. We've written before that trust isn't a feature you add later - it's just as true when the thing acting is an agent as when it's a payment.
What the swarm is allowed to do alone - and what it isn't
The freedom is in the drafting, not the delivery. On its own, the swarm can research, monitor, summarise, and generate as many drafts as it likes - none of that reaches anyone. What it cannot do alone is anything person-facing or irreversible: sending, posting, connecting, or committing us to something. Those cross the human gate. Rate limits, channel terms-of-service, consent and opt-out rules, and clear disclosure that a message is from a business are wired in as constraints the agents can't override - the same way a cashier can record a refund but only an approval releases the money.
This is the discipline we already apply everywhere else in the portfolio: assume the automated actor will get something wrong, and make sure the system can see it, bound it, and undo it. An agent is just a fast, tireless, literal new operator - and operators get scoped access, not the keys.
Where it plugs in
The swarm doesn't float in space. Approved engagements log straight into the portfolio tracker we're building - Helm - as targets, campaigns, and engagement history, so the marketing that happens and the record of it are the same thing. It runs on the stack we already use for production work: a queue-based orchestrator, the Claude API for the agents, and the Model Context Protocol to give them safe, scoped access to tools - the same agent-and-MCP muscle we described in building software agents can use.
Honest status
This is a roadmap piece, not a launch. The architecture is clear and the pieces are ones we've shipped before in other forms; the swarm itself is being built. I'm writing it down publicly for the same reason we publish everything else - because a direction you'll be held to is worth more than a demo you can hide. If it works, you'll notice Mpaukwu showing up more, more usefully, and never in a way that feels like a robot cornered you at a party.
The takeaway
The exciting part of marketing automation is that agents can work while you sleep. The important part is that a human still decides what reaches another human. Build the swarm for reach; keep the person on the send button for trust. Do both, and you get a studio that markets itself around the clock without becoming the thing everyone mutes.
At Mpaukwu, that's the plan: let the swarm do the tireless work, and keep me accountable for every word that lands in someone's inbox.
Mpaukwu Trading builds founder-led SaaS products, automation systems, and applied-AI platforms for African businesses. Read more from the Engineering Journal, see Helm, or start a project.