Building the product stopped being the hard part.
You can vibe code an MVP in a weekend now. Everyone can. Which means the entire game has shifted to one question: how do you get customers?
Cody Schneider went on Greg Isenberg's podcast and gave the clearest answer I've heard yet: marketing agents. Not a linear Zapier automation. Code running in the cloud, making decisions off your live business data, on a loop, while you sleep.
By the end of the episode he's described a system that researches your customers' pain points, generates ad creative, publishes it to Facebook, kills the losers after 2 to 3 days, and promotes the winners into a pool where they fight each other for budget.
Work that used to take an agency two weeks (and tens of thousands of dollars a month), set up by one founder in about 90 minutes.
Here's the full breakdown, with the source video so you can watch it yourself.
What a marketing agent actually is (and what it isn't)
Cody is blunt about this. He's talked to plenty of people who "talk huge game on the internet" about agents, and when you get into the details, there's nothing there.
His definition has three parts:
- → Unified data. The agent sees your whole pipeline in one place, not one channel at a time.
- → Autonomous decisions on a cadence. A thinking loop that runs on a schedule, not a one-shot prompt.
- → Hosted in the cloud. Real code on a server, not a workflow tool.
He's also clear about what he doesn't want: some fully autonomous AGI thing that "runs your entire marketing stack." Nobody has that, whatever their landing page says.
What he does have is an agent that entirely runs Facebook ads for real companies. It researches the target customer's pain points, makes on-brand creative (statics and AI avatar UGC), publishes into the ad account, turns off losers, scales winners, then studies what worked to make the next batch better.
A narrower claim than the AGI crowd makes. But he can actually show it running.
The startup idea Cody keeps pitching: AI for WordPress
Before the agent playbook, the episode opens with a business idea Cody has pitched to three YC founders looking for a pivot. All three said no. He still thinks they're wrong.
WordPress powers 43% of the websites indexed by Google. Almost half the internet. And nobody is building AI-first products for it.
The pitch: Lovable, but for WordPress. Let people vibe code their site on top of the platform they're already on, bundle the plugins they currently duct-tape together (forms, CRM, SEO), and sell tokens starting at $29 a month.
Greg pushed it further. At 19 he was a partner in an agency that migrated big sites onto WordPress, officially partnered with WordPress itself. Time.com, TechCrunch, that tier. He watched absurd money flow through the plugin ecosystem.
His version of the play: find plugins with validated demand and no AI component, then build the AI-first version of each.
- → Yoast shows you red and green dots and tells you to fix things yourself. An agent just fixes them: writes the meta, restructures content, adds internal links.
- → WPForms is a static form. Make one that talks back, qualifies the lead, answers questions.
- → WooCommerce gets an AI storekeeper: product descriptions and abandoned cart flows on autopilot.
Proven demand, people already paying, and you make it 10x better with an agent. The product is the easy half. The rest of the episode is about the hard half: distribution.
Andromeda changed Facebook ads: your creative is the targeting now
This was the part that made me sit up.
Facebook's new ad algorithm, Andromeda, killed interest-based targeting. You don't pick "people interested in WordPress" anymore. The AI reads your creative (the image, the text, the video script) and your landing page, and decides who sees the ad.
So the targeting lives inside the ad itself. Write a script like "I was paying my agency $1,000 a month to maintain my WordPress site and they were never responsive. Then I found this plugin that lets me chat with AI to change my website." Facebook figures out who has that exact problem and finds them.
Cody says something wild here: Facebook has turned into the best B2B ads channel that exists right now. His team runs ads so obscure that maybe 10 people in the US have the problem, and Facebook finds those 10 people.
Greg's take on why founders fail at this: they try a few ads, nothing happens, they quit. The fix is running the same ad with the positioning changed 10, 15, 20 times. Sometimes the angle you'd never bet on ("anti-Yoast" instead of "better Yoast") is the one that converts. You can't predict it. You can only test it, and paid ads give you a clear answer within 48 hours.
As Cody puts it: we're not Don Draper. You don't impose your idea on the market. You put a thousand creatives in front of it and let it tell you what it wants.
The stack: a data warehouse, a pipeline, and an agent in the cloud
Here's where most "agent" content stays vague and Cody doesn't.
The infrastructure is three pieces, all self-hostable, all open source or cheap:
- → Airbyte for the data pipeline. Pre-built connectors that pull Facebook Ads, Google Analytics, PostHog, HubSpot, and Stripe data into one place.
- → ClickHouse as the data warehouse. Every data source in context with every other one.
- → Any cloud for hosting the agent: Heroku, Railway, whatever. You don't need a Mac mini running in your closet.
Why the warehouse matters: to run Facebook ads well, you need to connect the specific ad someone clicked to the revenue that came out the other side. Ads data alone can't tell you that. Ads plus analytics plus CRM plus Stripe, in one queryable place, can.
A side effect I love: once the warehouse exists, you get conversational analytics for free. Cody's example is a business owner literally asking, "We're having trouble hitting payroll this month, what's going wrong?" and getting "your accounts receivable is off" back. Your marketing data becomes something you talk to.
One warning worth its own paragraph. People keep getting their Facebook ads accounts banned and blaming the agent. The agent isn't the reason. They spammed the Marketing API trying to pull millions of rows of data, which violates Facebook's terms. Use the API for writes only: publishing ads, turning them off, scaling them. Read your performance data from the warehouse instead.
10 ads a day, a 3-day window, and a winners pool
The numbers from a company Cody's team runs this for:
- → 2 ad sets per day, 5 ads per set. 10 fresh ads daily, generated and uploaded automatically.
- → Each batch runs 2 to 3 days to get initial signal.
- → The agent pulls performance from the warehouse, turns off the worst performers, and promotes winners into a pool where they compete for the available budget.
Creative comes from research, not vibes. The agent scrapes Reddit (via Perplexity) for the pain points real people complain about, ranks them by how often they come up, and builds ads around the top three. Statics get generated with Nano Banana, checked by a vision model against brand style guides (right fonts, right colors, readable text). Video is HeyGen avatar UGC, with Seedance experiments coming.
Then the loop closes. Every prompt and script that produced an ad goes into a database, so the agent can study which inputs made winners and generate more like them.
Greg's framing for why this matters: paid ads are the only system where you put $1 in and $5 can come out the other side, and you can keep feeding the machine. Most founders never get there because the old version required an agency and tens of thousands a month. This version requires a weekend of setup.
The entropy problem nobody talks about
Set this system up and day one feels magical. Day three, the ads start looking the same. Day five, performance sags.
Cody calls it entropy: the agent gets stuck thinking the same way. And he points out that anyone who says "you set it up and it just works" has never set one up.
Two fixes, both about injecting new DNA into the system:
- → Pull competitor ads from the Facebook Ads Library and feed them in as fresh source material.
- → Mine YouTube and podcast transcripts in your niche for insights, and build ads off those. There are hundreds of channels dedicated to almost any category.
There's also Viral Loop, a tool with an API that scrapes trending TikTok content, so the agent can see which formats are blowing up in your category this week and adapt.
The mental model shift: marketing stopped being campaigns you start and stop. It's a continuous loop reacting to a market that moves 10x faster than it did 20 years ago (Cody's comparison is how fast fashion trends cycle now versus the early 2000s). You want a system that notices the shift before you do.
What this means for you
My favorite moment in the episode is when Cody describes marketers asking him "is my job secure?" His answer: "I don't know. Can you do this?"
The job becomes agent jockey. You take the domain knowledge in your head and encode it into a system that executes it a thousand times faster than you could. Cody says he's not technical, and he can set this up. Anyone with Claude Code and this episode's transcript can walk themselves through it.
I've written before about why AI agents are the new SaaS and why managing agents is the most valuable skill of 2026. This episode is what those ideas look like deployed against the single hardest problem every bootstrapped founder has: getting customers without a budget for an agency.
You don't need the full system on day one. Start with one loop. Warehouse your ads and Stripe data, generate 10 ads a week from real Reddit pain points, kill losers every 3 days. Then let it compound.
FAQ
What is a marketing agent?
A marketing agent is code hosted in the cloud that makes marketing decisions autonomously off your live business data. Cody Schneider's definition has three requirements: unified data across your whole pipeline (ads, analytics, CRM, payments), a decision loop that runs on a cadence, and cloud hosting. A linear automation that fires the same steps every time doesn't qualify.
Will using an agent get my Facebook ads account banned?
Not if you use the Marketing API correctly. The bans people blame on agents come from spamming the API with massive data pulls, which violates Facebook's terms of service. Use the API only to write: publish ads, pause them, adjust budgets. Read performance data from your own warehouse, fed by a pipeline tool like Airbyte.
What tools do you need to build one?
The stack from the episode: Airbyte for the data pipeline, ClickHouse for the warehouse (both open source and self-hostable), Heroku or Railway to host the agent, Perplexity for pain point research, Nano Banana for static creative, and HeyGen for AI avatar video. Total cost is a fraction of one month of agency retainer.
How is this different from hiring an agency?
An agency running creative testing at this scale costs tens of thousands of dollars a month, and historically getting 100 ads made and published took around two weeks. The agent system produces 10 fresh ads a day, reacts to performance data within 48 hours, and runs continuously. The setup Cody describes takes a semi-technical founder about 90 minutes to stand up in basic form.
Steal playbooks like this every week
I interview bootstrapped SaaS founders making $100K to $10M a year and pull out exactly how they get customers, price, and grow. No VC noise, just operators showing their numbers.
