Marketing Agents: What Actually Counts as One

Cody Schneider draws a hard line between an agent that owns your data and an automation that just fires webhooks

There is a word doing an enormous amount of heavy lifting in software right now, and that word is agent. It is on landing pages, in funding decks, in the release notes of tools that, six months ago, were perfectly happy calling themselves automations. The label spread faster than the meaning did.

So it is genuinely refreshing when someone stops and draws a line. On a recent episode of Greg Isenberg’s podcast, marketer and operator Cody Schneider does exactly that. He sets a bar for what earns the name — and by his own admission, most of what ships under the agent banner does not clear it.

The three-part test for a real agent

Schneider’s definition has three parts, and all three have to be true at once.

• It owns unified data. Not a slice of one platform — the whole pipeline, in one place, live.
• It runs on a cadence. It makes decisions on its own schedule, not because a human clicked something.
• It reads its own results back. There is a thinking loop: the output of yesterday’s decision becomes the input to today’s.

Miss the third one and you have a very sophisticated cron job. Miss the first and the agent is making confident decisions on a partial view of reality, which is arguably worse than making no decisions at all.

What I like about this framing is that it is testable. You can hold any product up against it and get a yes or a no in about ninety seconds. A linear automation workflow — trigger fires, webhook posts, spreadsheet updates — fails on cadence and feedback. It does a thing. It does not learn whether the thing worked.

An agent that cannot read its own results is a very sophisticated cron job.

The stack is shorter than you think

Here is the part that should be encouraging to anyone who has been putting this off. When Schneider lists the infrastructure behind a working marketing agent, it comes down to three pieces.

Pipeline, warehouse, hosting

The pipeline is Airbyte — the thing that pulls data out of the places it lives. The warehouse is ClickHouse — the place it all lands and gets queried. And the code runs in the cloud on something like Heroku or Railway. That is the list.

In the ads example, Airbyte pipes Facebook ads data, Google Analytics, CRM records and Stripe revenue into ClickHouse. Suddenly the agent is not looking at a Facebook dashboard that thinks every click is a conversion. It is looking at spend on one side and actual money on the other, in the same table.

That is the unlock, and it is not glamorous. Most of the intelligence in a marketing agent comes from the fact that it can finally see the whole board. If you have been reading up on how AI infrastructure is being discussed across the technology and AI channels, you will recognise the pattern: the model is rarely the bottleneck. The data plumbing is.

WordPress as the sandbox: a $10M idea in plain sight

The concrete business Schneider and Isenberg use to make this tangible is an AI-first product built on WordPress. And the opening statistic is hard to argue with: WordPress powers 43% of all indexed websites.

Think about what that means. Nearly half the web runs on a stack where the standard experience is stitching together a dozen plugins — forms here, CRM there, security somewhere else — each with its own billing, its own update cycle, its own settings page nobody understands.

The pitch is a vibe-coding layer for WordPress that bundles those plugins into one package, sold on tokens, with a $29-per-month base tier. Schneider puts a number on it: a $10 million a year company inside roughly 24 months, for a team that executes and positions it well.

That is a bold claim and worth treating as one. But the underlying logic is sound, and it generalises.

Rebuild the plugins that already proved demand

Isenberg brings agency experience to this — migrating time.com and techcrunch.com onto WordPress — and notes the genuinely absurd amount of money moving through the plugin ecosystem.

That produces a repeatable play: find plugins with validated demand, then ship the AI-first version of each.

• Instead of Yoast’s red and green dots telling you your meta description is too long, an agent that writes the meta and the internal links.
• Instead of a static contact form, a conversational form that qualifies the lead while it is talking to them.
• Instead of a WooCommerce admin panel, a storekeeper that writes product descriptions and runs abandoned cart flows.

The demand is already validated. Somebody else spent a decade proving people will pay for this category. You are not guessing whether the market exists — you are betting that the AI-native version is enough better to make people switch. That is a much smaller bet, and it is the kind of idea that turns up constantly in the startup and entrepreneurship conversation.

Why the ad copy is now the targeting

The most consequential shift in the episode is about Meta. Facebook’s Andromeda algorithm reads your creative and your landing page — which means the ad copy now carries the targeting.

For a decade, performance marketing was an exercise in audience construction: interests, lookalikes, exclusions, layered stacks of demographic guesses. That skill has been quietly commoditised. The algorithm figures out who to show it to. Your job is to give it something worth reading.

The algorithm decides who sees it. You decide whether it says anything worth seeing.

Which reframes the entire job. If creative is targeting, then the constraint is how many distinct angles you can produce and test — and that is a volume problem, which is exactly the kind of problem an agent is good at.

Where the angles come from

Schneider’s research step is unfussy and easy to copy. He runs Perplexity against Reddit to pull the real pain points and desired outcomes of WordPress site owners, then ranks them by how often people bring them up.

The top three come back as: the plugin layer, performance and speed, and security and maintenance. Those become the ad angles. Not invented positioning — the words people already use when they complain.

Two creative pipelines

Statics come out of Kai AI with Google Nano Banana, seeded with an example ad from a competitor or from a tangential industry. A vision model reviews every output against brand style guides — fonts, colours, whether the text is actually readable at thumbnail size. That review step is the difference between volume and slop.

Video runs on HeyGen for avatar UGC today, with Seedance coming up fast. Clips cap out around nine seconds, so anything longer needs frame stitching — a real constraint worth knowing before you plan a thirty-second spot.

Fresh inputs keep this from going stale. Competitor ad libraries, YouTube transcripts, podcast transcripts — all of it feeds back in so the creative keeps varying over time instead of converging on one template the agent likes. Schneider calls this solving for entropy, and it is the failure mode nobody warns you about: an agent left alone will happily produce the same ad five hundred times.

Let the market pick the winner

The strategy that falls out of all this is almost aggressively simple. Publish through the Facebook Marketing API, run a thousand angles, and read the market’s verdict inside 48 hours. Kill the losers. Promote the winners.

No committee debating which concept is on-brand. No two-week creative review. The market answers faster and more honestly than any internal opinion, and it answers in the only currency that matters.

There is a discipline hiding in here that is worth naming, because it is the same one that separates good operators from busy ones across the sales and marketing landscape: you have to actually kill things. Most teams are far better at launching tests than at ending them. An agent that runs on a cadence does not get attached to the ad it made on Tuesday.

What to take away from this

The honest read on this episode is that the hard part was never the AI. It is the plumbing, the cadence, and the willingness to let results overrule taste.

If you are evaluating a tool that calls itself an agent, run the three-part test. Does it own unified data across your whole pipeline? Does it decide on a cadence without you? Does it read its own results back and change what it does next? Two out of three is an automation with good marketing.

If you are building, the sequence Schneider describes is genuinely followable: get the data into one warehouse first, because everything downstream depends on it. Find the angles in the language your customers already use, not in a positioning workshop. Produce creative at volume with a quality gate on the output. Then let 48 hours of spend tell you what works.

And if you are hunting for a product idea, the WordPress example is a template more than a tip. Look for a category where demand is already proven and the incumbent interface is a dashboard full of settings. That is where the AI-first version has room to win — and it is a lens that applies well beyond plugins, as anyone tracking how growth-stage businesses are rethinking their tooling will have noticed.

The biggest insight, though, is smaller and more useful than any of that. An agent is not a model with a personality. It is a feedback loop with access to the truth. Get the truth into one place, give it a schedule, and let it learn from what it did yesterday. Everything else is a wrapper.

▶

Sources

• Marketing Agents Are Too Good Now — Greg Isenberg with Cody Schneider — Full episode (approx. 38 min). Key sections: defining a marketing agent (01:54), the AI-for-WordPress idea (04:00), the state of Meta ads (09:55), creative pipelines (15:23), data pipeline and warehouse (17:25), solving for entropy (25:51), letting the market pick the winner (28:01).