Somewhere in Purple this week, someone shipped something genuinely clever. A tool that quietly kills an hour of manual work a day. A pipeline that catches a class of bug before it reaches production. And a month from now, almost nobody will know it exists.

Not because it isn't good. Because telling the story of it well is a whole job. It needs someone to understand the thing, find the angle, write the article, make the artwork, record a version people can listen to, build a page, and write the launch post — and the small team that built it doesn't have a marketer, and never will. So the work is real and the story never gets told. It dies in a repo. Multiply that by an innovation function shipping a steady stream of internal tools, and you have a company that keeps getting better at things it doesn't know it can now do.

This is the story of the team we built to stop that happening. The twist is that the team is made of files and agents — and it made everything else you've been reading.

We built a marketing team as software

The idea is almost embarrassingly literal. If the problem is "we don't have a standing content team," then build one — as software. Not a chatbot you prompt from scratch each time and that forgets everything the moment you close the tab, but a standing team that lives as files in a workspace, so it remembers what it did last time and plans what's next across sessions.

That's the important word: standing. The Marketing Engine has memory. Every job it does starts by reading a running context log and a backlog, and ends by updating them. Decisions get written down and honoured later. A house style persists. So the tenth episode doesn't start from zero — it starts from everything the first nine taught it. It behaves less like a clever one-off and more like a small publication that grows.

How it works: five stages, one show-runner

Point the engine at a product and it runs a five-stage pipeline — MINE, PANEL, PRODUCE, REVIEW, DEPLOY — with a show-runner in charge and specialists convened for each stage. There are twelve roles in all, in three groups, coordinated by an Orchestrator: a lead agent whose only job is to decide what stage we're in, bring in the right people, force a decision when the group stalls, and keep the memory honest. It protects the editor's time by escalating only the handful of things a human actually has to decide.

First comes research. A Story Miner reads the product's repo, its commit history and its docs, drafts a "story bank" — the single source of truth for that episode — and interviews a human to pull out the hook and the framing that no document contains.

Then the panel argues. A Brand Strategist, a Copywriter, a Technical Translator, an Audience-Exec stand-in, and a Skeptic debate the positioning and hand up a written brief. This is where the angle gets fought over before a single sentence of the article is drafted — so production isn't guesswork, it's execution against a decision.

Then the production crew builds it. A Blog Author writes the full article to the brief. An Art Director generates the imagery — in two house looks, "Aurora" and "Gazette," so the series has a consistent visual identity. A narrator turns the finished article into a "listen to this" audio version. A Video Scriptwriter writes the teaser, and a Social Promoter writes the Slack launch post. Finally, one command deploys the whole series as a single static site, and a shared manifest keeps the hub, every sidebar and every audio player in lockstep, so adding one episode updates the entire publication.

The part that makes it trustworthy

Here's the turn, and it's the reason this episode exists. Anyone can wire up "AI, write me some marketing." The result is usually confident, fluent, and quietly full of things that aren't true — an invented statistic here, a fabricated customer quote there. For internal storytelling that leaders will act on, that's worse than useless.

So the most important thing the Marketing Engine has isn't its writing. It's its honesty engine — three mechanisms whose entire purpose is to stop it lying, even by accident.

The first is the Skeptic. It's a named member of the panel with veto power, and nothing is allowed to be called "done" until it signs off. Its job is to block anything unproven — an unverified number, an overclaim, a metric nobody actually measured. That's why this very episode has no impressive statistic about itself: the Skeptic won't let the engine make one up, not even in its own favour.

The second is locked memory. Once a human settles an angle or a framing — "keep this qualitative," "don't name that system," "this claim needs sign-off" — that decision is written into memory and treated as a hard constraint on every future run. The engine can't quietly drift back to a version someone already rejected.

The third is the assets-needed ledger. When the engine hits something it genuinely can't get or verify on its own — a screenshot behind a VPN, a real quote, a number that needs a human's sign-off — it doesn't paper over the gap. It writes down exactly what it needs and asks. It never invents a quote, and it never goes hunting for secrets to get around a wall.

Put together, these do something specific: the machine does the tedious ninety percent — the research, the drafting, the art, the narration, the build, the deploy — and a human keeps the final call on the two things a machine shouldn't own alone: what's true and what the angle is. It's the same human-in-the-loop shape you saw in the Nexus episode, where a person still approves what the AI suggests. Here, the human is the editor, and the agents are the newsroom.

The proof is the thing you're holding

The easiest way to be skeptical of a marketing machine is to assume its own marketing is the one thing it's best at faking. Fair. So here's the proof that isn't a claim: this series.

Six episodes have already shipped through this pipeline — the Innovation Department, the Architecture Workspace, the Living Map, Nexus, Vantage, and the Investec deposit feed — each mined, argued, written, illustrated, narrated, built and deployed the same way. The heroes at the top of each one are generated images. The "listen to this article" audio is synthetic narration. The words were drafted by an agent working to a brief a panel fought over. And this page — the one you're reading right now, about the engine — was produced by the exact pipeline it's describing.

That's the honest flex. Not "AI can write marketing." It obviously can. The point is that a small, file-based team of agents, kept honest by a Skeptic and a human editor, can run an entire editorial function — and you can't easily tell which part a person did, except for the parts we deliberately reserved for one.

One honest word

The engine is real and in use, but it's a capability, not a finished product, and we'll say so plainly. A human still locks the angle and still supplies the real quotes — including this one, which is why there's a space held open for it rather than a convincing sentence we made up.

It's still maturing. There are things it can't do without a person, on purpose. But the direction is clear, and the compounding is already visible: each episode starts further ahead than the last.

Who this is for, and where it fits

For the innovation teams, it means the thing you built finally gets told — to a consistent standard, without you having to become a marketer for a week.

For everyone at Purple, it means a steady, honest drumbeat of what's being built and what it now makes possible, instead of finding out by accident months later.

For leaders, it's a browsable, growing record of the whole innovation portfolio — each piece cited, checkable, and honest about what's proven versus what's promising.

And in the bigger picture, this is the storytelling layer of the innovation operating system. The other episodes are about how Purple knows what it runs, designs what's next, and builds it with AI in the room. This one is about how it tells that story — automatically, honestly, and at a pace no single marketer could keep.

What's next

Tighten the loop. More episodes, produced faster, from a memory that gets richer each time. A closer hand-off between the product teams and the engine, so a shipped tool can nominate itself for the series. And, deliberately, the one thing that stays exactly where it is: a human keeping the final call on what's true and what the angle is. The machine is allowed to do almost everything. It is not allowed to decide, on its own, what's worth believing.