How to Build an AI Marketing Team That Runs Itself
A HubSpot podcast interview that turns into a live Claude Code build session, wiring a Facebook ads agent, a Google Ads agent, an SEO agent, and a cold-email workflow into one voice-driven marketing stack.
Posted
3 days ago
Duration
Format
Interview
educational
Views
8.4K
122 likes
57 · 43
Big Idea
The argument in one line.
A coding agent given real API access to your ad platforms, CMS, and data warehouse can research, create, publish, and optimize marketing campaigns at a volume and speed no human team can match, turning marketing itself into a codebase.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You run paid ads (Meta or Google) and want to understand how agencies are using AI agents to cut cost-per-lead in weeks, not quarters.
You already use Claude Code or a similar coding agent for engineering work and want to see it pointed at marketing tasks instead.
You're a marketer or founder deciding whether to build custom AI-agent infrastructure versus buying a tool like n8n or Zapier.
You lead a small marketing team and are wondering whether headcount or tooling is the better next hire.
SKIP IF…
You want a no-code, drag-and-drop automation tutorial — this is explicitly the opposite pitch ("go straight to code").
You're looking for Facebook/Google ads copywriting tips rather than the infrastructure and agent-orchestration layer underneath them.
TL;DR
The full version, fast.
Cody Schneider argues that AI targeting has gotten so good that the bottleneck in paid ads is no longer data or targeting, it's creative volume and idea quality: campaigns that used to run 5 ads now need 50-500. His agency treats every ad as a SKU-tagged JSON blob stored in Postgres, connects performance data back to each asset through a data warehouse, and lets a coding agent rewrite ad sets, keywords, and landing pages daily. He then builds four agents live in Claude Code by voice: a Meta ads creative agent, a Google Ads keyword agent, an SEO agent that researches/writes/publishes articles, and a cold-email workflow from LinkedIn engagers. The throughline: give a coding agent real API access plus a documented "taste profile" of your brand, and it outperforms ad-hoc human execution at scale.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Cold open: Cody previews the Facebook ad results (cost per lead $82 to $15) and states the episode's thesis, that a coding agent with the right context is now the best marketer on the team.
00:50 – 02:25
02 · What an AI agent actually is (not God in a box)
Host intro of Cody Schneider and Graphed. Cody defines an agent as software with a thinking loop running on a live data stream, pushing back on the idea of an autonomous 'God in a box' running an ad account unsupervised.
02:25 – 03:10
03 · Why ads now need 10x the creative
Because AI targeting is now so effective, Cody argues the constraint moved from data/targeting to creative volume and idea quality: campaigns that used to run 5 ads now need 50 to 500.
03:10 – 05:35
04 · All marketing is code: 10 ads a day learning loops
Graphed launches roughly 10 new ads per day rather than dumping hundreds at once, treating each static image as a JSON blob with a SKU-style ID stored in Postgres so performance data can be tied back to the exact asset.
05:35 – 07:05
05 · The results, and why your coding agent is your best marketer
A private-equity portfolio company (250 locations) cut Google Ads cost-per-lead from $70 to $35 and Facebook cost-per-lead from $80 to $15 in four weeks on the same budget, driven by creative volume and automated negative-match keyword pruning.
07:05 – 09:05
06 · The marketing engineering stack your agent needs
Cody lists the infrastructure a coding agent needs to do real marketing work: data pipeline, data warehouse, cloud server, media storage, an agent database, cron jobs, app authentication, shareable links, a Git manager, and a tailored API gateway to tools like Nano Banana, Apollo, Seed Dance, and Apify.
09:05 – 12:35
07 · Giving up creative control: taste profiles and human review
Addressing marketers' discomfort with agents publishing creative they didn't personally make, Cody introduces the 'taste profile' concept from his book Loop and describes human-in-the-loop review (including Slack approval gates) as the bridge to full automation, especially in regulated industries.
12:35 – 14:15
08 · The Google Ads agent that rewrites its own code
A case study: a startup's Google Ads cost-per-lead fell from $1,100 to $250 over four months as an agent that rewrites its own deterministic logic daily handled keyword research, ad sets, ad copy, and landing pages.
14:15 – 17:25
09 · Live in Claude Code: building a Meta ads agent by voice
Cody dictates a live prompt in Claude Code asking it to scrape Graphed's own site style guide and generate 1080x1080 static ad creative via the Nano Banana API, framing voice-driven prompting as the emerging default way to direct a coding agent.
17:25 – 19:15
10 · SEO plugin, and why Facebook bans agent-run ad accounts
Cody installs an SEO plugin through the Graphed CLI and digresses on why agent-run Facebook ad accounts get banned: agents spamming read requests against the API, not the writes, which is why analysis should run from a data warehouse and the live API should be reserved for writes.
19:15 – 22:00
11 · What we built: creative, keywords, SEO, cold email
Recap of the parallel builds: Nano Banana creative generated from the scraped site palette, Google Ads keyword research via a data-for-SEO API plus Google Search Console, an SEO article drafted and published to a Strapi CMS, and a cold-email workflow pulling LinkedIn post engagers into a waterfall enrichment and validation pipeline feeding Instantly.
22:00 – 24:20
12 · Friends don't let friends do n8n
Cody argues coding agents have made drag-and-drop tools like n8n and Zapier unnecessary for this kind of work, since you can simply ask the agent to visualize the system it built, and reframes 'agents' as just threads inside a coding assistant doing tasks against a properly wired-up stack.
24:20 – 26:05
13 · Is the CMO role dying?
The hosts and guest debate whether the traditional CMO role survives: one person with the right tooling and data access can now do the work of 10 to 100 people, but organizations still need someone to manage and orchestrate at scale.
26:05 – 28:14
14 · Where to start: give your coding agent API keys
Closing advice: store your platform API keys locally, hand them to whatever coding agent you already use, and try to get it to do a real task you'd normally do by hand, then turn a working manual session into a repeatable, scheduled agent.
Atomic Insights
Lines worth screenshotting.
Cost per lead on Facebook dropped from $82 to $15 (seven-day trailing average) for one client after switching to AI-agent-driven creative production.
A separate portfolio company took Google Ads cost-per-lead from $70 to $35 in four weeks, and Facebook cost-per-lead from $80 to $15 in the same window, on the same budget.
One startup's Google Ads agent, which rewrites its own optimization logic daily, cut cost-per-lead from $1,100 to $250 over four months.
Because AI ad targeting is now this good, the constraint shifted from data and audience targeting to creative volume and idea quality: campaigns that ran 5 ads now need 50 to 500.
An 'agent' is just software with a thinking loop running on a live data stream, not an autonomous black box making unsupervised decisions.
The agency treats every ad creative as a JSON blob with its own SKU-style ID, stored in Postgres, so ad-platform performance data can be joined back to the exact asset that produced it.
Launching creative in small daily batches (about 10 new ads a day, 300 a month) instead of dumping 300 ads at once avoids getting stuck in each ad platform's learning phase.
A 'taste profile' is a single file combining brand identity, customer data, and emotional triggers that lets an agent generate on-brand creative without a human designing every asset.
Facebook bans agent-run ad accounts mainly because agents spam read requests against the API, not because of the writes; the fix is doing all analysis from a data warehouse and reserving API calls for writes only.
Regulated industries can encode allowed and disallowed language directly into the agent's system so compliance is enforced automatically rather than reviewed after the fact.
The stack a coding agent needs to do real marketing work: a data pipeline, data warehouse, cloud server, media storage, a database for agent state, cron-based recurring jobs, app authentication, shareable links, a Git manager, and a unified API gateway to tools like image generation, LinkedIn scraping, and SEO data.
The workflow described skips MCP servers in favor of a single NPM package that catalogs roughly 230 tool endpoints the coding agent can call directly.
Voice-driven prompting (dictating the task instead of typing it) is presented as an increasingly normal way to direct a coding agent through a multi-step build.
The speakers argue that building software was never the real barrier; understanding the problem deeply enough to build the right thing, and sustaining the upkeep afterward, still is.
One two-person marketing team at a 15-year-old company chose to buy agent tooling instead of hiring, turning each person into what the guest calls a '100x employee.'
Takeaway
Your coding agent can run the whole marketing stack
AGENT PLAYBOOK
Give a coding agent real API access to your ad platforms, CMS, and a data warehouse instead of just prompts, and it can out-produce a human team on creative volume, keyword pruning, and publishing speed.
01$82 to $15 cost per lead with AI agents
A coding agent given real API access to your ad platforms, CMS, and data warehouse can research, create, publish, and optimize campaigns faster than a human team.
02What an AI agent actually is (not God in a box)
'Agent' just means software with a decision loop reacting to a live data feed, not an unsupervised black box; use the LLM only where inference is genuinely needed.
03Why ads now need 10x the creative
AI targeting has gotten good enough that the real bottleneck moved from audience data to creative volume: plan for 10x the ad variations you used to run.
04All marketing is code: 10 ads a day learning loops
Launch new creative in small daily batches (roughly 10 a day) instead of dumping hundreds at once, so each ad platform's learning phase doesn't stall.
Tag every piece of creative with a stable ID and store it in a database so ad-platform performance can be joined back to the exact asset that produced it.
05The results, and why your coding agent is your best marketer
Automated negative-match keyword pruning plus higher creative volume, not fancier targeting, is what moved cost-per-lead in the case studies discussed.
06The marketing engineering stack your agent needs
The infrastructure a marketing agent needs is boring but specific: a data pipeline, data warehouse, cloud server, media storage, an agent database, cron jobs, auth, and a unified API gateway to your tools.
07Giving up creative control: taste profiles and human review
Before you let an agent publish anything, write down a 'taste profile': your brand voice, fonts, colors, and customer emotional triggers in one file it can reference.
Add a human approval gate (even something as simple as a Slack message) for new creative until you've built enough trust in the outputs, especially in regulated industries.
08The Google Ads agent that rewrites its own code
An agent that rewrites its own optimization logic daily based on live performance data can keep improving a funnel without a human re-touching it each time.
09Live in Claude Code: building a Meta ads agent by voice
Dictating a build to a coding agent in plain spoken language, rather than typing a careful prompt, is becoming a normal way to direct multi-step marketing builds.
10SEO plugin, and why Facebook bans agent-run ad accounts
If your agent is getting banned from an ad platform, the likely cause is spamming read requests against the live API; do analysis from a data warehouse and reserve API calls for writes.
11What we built: creative, keywords, SEO, cold email
A single documented tool stack (creative generation, keyword data, a CMS, an email verifier, and a sending tool) is enough for a coding agent to run ads, SEO, and cold email in parallel.
12Friends don't let friends do n8n
Before reaching for a drag-and-drop automation tool, try asking your coding agent to build and then visualize the workflow directly in code.
13Is the CMO role dying?
One person with the right data access and tooling can now cover work that used to require a small team, which changes headcount math more than it eliminates leadership.
14Where to start: give your coding agent API keys
The actual bottleneck was never whether you can build something with agentic coding, it's whether you understand the problem well enough to build the right thing and maintain it.
Start small and concrete: put your platform API keys in a local environment file and have your coding agent attempt one real task you'd normally do by hand.
Glossary
Terms worth knowing.
Taste profile
A single file capturing a brand's identity, customer data, and emotional triggers so an AI agent can generate on-brand creative and copy without a human designing every piece.
GTM engineering
Treating go-to-market work like software engineering: building systems, monitoring them, finding what works, and automating and auditing the workflow rather than running it manually.
Negative match keyword
A Google Ads setting that blocks a specific search term from ever triggering an ad again, used to cut spend on traffic with no real purchase intent.
Waterfall enrichment
Trying multiple data or email-finding providers in sequence until one successfully returns and validates a contact's information.
Forward deployed engineering
An engagement model where a vendor's team works hands-on inside a client's own data and systems to build and ship something custom, rather than selling off-the-shelf software.
Anti-token-maxing / outcome-maxing
A design philosophy that minimizes how much work is handed to an LLM and maximizes how much runs as plain deterministic software, calling on the model only where inference is actually needed.
n8n
An open-source, drag-and-drop workflow automation tool; used here as the example of the visual-builder approach the guest argues coding agents have made unnecessary.
“Your coding agent is now the smartest marketer, the best marketer in the world, if you give it all the context that it needs to do its job.”
the episode's one-line thesis, quotable on its own→ TikTok hook↗ Tweet quote
01:00
“Oh, we put God in a box and let it run a Google Ads account. Like that's... you don't need God in a box to run a Google Ads account.”
punchy pushback on agent hype, sets up the real definition→ IG reel cold open↗ Tweet quote
06:15
“It's almost like this quant, like function that's on top of this technology... this is a software engineer running this account, right? It's just the velocity that can happen at.”
“Building stuff was never the barrier. It was: do I really understand the problem to build the right thing?”
reframes the whole 'anyone can build now' narrative→ newsletter pull-quote↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphor
I want to make some Facebook ads for the ICP that is basically trying to connect their marketing data to cloud code. You need really good creative and you need 10 times the amount of creative that you used to. On Facebook, we got their cost per lead down from $82 down to $15 is the average, like seven day trailing.
Your coding agent is now the smartest marketer, the best marketer in the world, if you give it all the context that it needs to do its job.
Cody, at graph .com, you are out there building in the marketing world, kind of the frontier of marketing and AI. And I think you've learned a lot. And on today's show, you're going to break down what are the key agent use cases in marketing right now.
And we're going to build and deploy some agents live. So before we get into the agent building, maybe walk us through. You're talking with tons of marketers out there.
Where are they finding success with AI and agents today? Yeah, just to kind of zoom out, like what is an agent, right? It is software with a thinking loop, with a live data stream that it's making decisions on top of, right?
Like at its core, fundamentally, that's what it is. I think people think about an agent, they're like, oh, we put God in a box and let it run a Google Ads account. Like that's, you don't need God in a box to run a Google Ads account, right?
Like you, it's a very - You have your ad strategy and an agent can help you manage and do that on a proactive basis, right? Exactly, right. And so actually what's more effective is I make - Custom software that's running a process that, again, maybe an LLM is like a part of that.
That's an agentic part of it, right? So for example, how we use it for like a Google Ads agent is we'll look at all the search terms that are coming up and we'll basically be like, okay, we're going to scale this one to 10. Is this search term have basically like the search intent that's related to the product?
If no, I'm going to, you know, basically it's a binary. I do a negative match on that keyword. I'm never going to spend on that keyword again, right?
You use the LLMs, you use inference, right? Only when you need that inference. I'm trying to run as much as I can as software because that's the cheapest way for me to do this.
And we're kind of like against this, like as a company, like our kind of whole thing is like, we're anti -token maxing, right? And that's what we're all about, outcome maxing. It's like, you're just trying to get to the best result.
And that's a blend of tokens. And then it's a blend of deterministic, traditional software. Exactly.
There's so many people just like building stuff that has no value in the world just because they can. And that's like the wrong way to do it. My observation, and see if you agree or disagree, is that the advertising world has changed most dramatically, probably any part of marketing outside of maybe search in the last couple of years.
And the big change that I don't think people fully come to reckoning with is that all the advanced AI targeting is now so good that now you need money and idea. You don't need the data and the targeting nearly as much as you used to. But while you need really good creative and you need 10 times the amount of creative that you used to.
And if you used to run a campaign with five ads, you probably now need to run a campaign with 50 ads, 500 ads. And that's what I'm seeing of the people who are really doing this at like the best, best level. Is that what you're observing?
Are you finding something different? No, we're finding the exact same thing. I think the only difference is that we're not launching simultaneously.
300 ads, right? Say, for example, we're going to launch 10 ads per day and we're doing it on that cadence so that we have these learning loops that occur where it's basically like, okay, all marketing is now code, right? Say you're making static images, right?
Those static images are just a JSON blob that got sent to Nanobanana or ChatGPT image 2 .5, right? So I can store that code in a database, like a Postgres database, and I can give each static image its own unique ID, like a SKU number, right? So when I upload that, you know, content into an ad set within Facebook ads, I can have basically a log of all of the information that was, I can have, what is the pain point desired outcome that this is based off of?
Did I pull this from a competitor ad library? And that was like the source material. Okay, here's the JSON blob.
Now I can connect that to my data. If I use a data pipeline data warehouse, I can actually connect my Facebook ads data to that specific ad. And then when I give that to my coding agent, It now can understand, OK, here is the type of content that's actually creating like outcomes.
Let's go in a learning loop. So every 10, you know, again, we're doing 10 new pieces of creative per day. We can go through two to three cycles per week on top of that.
That turns into those 300 ads per month. But just going and dumping 300 ads, what we found, unless you have a massive budget, it just ends up you get stuck in this learning phase, depending on the company type. And again, if you're if you're B2B or if you're consumer, there's a lot of different variables here.
If you're prosumer, et cetera. But. What we found is by going through these loops, you can basically figure out, we call it like a persona matrix.
I can figure out what are the emotional triggers? What are the situations that they're in? And what is the market actually responding to?
And then basically like, I think about it as like a field of view. I have a wide field of view and then it scopes into like, what is that thing? It's the exact same thing that we were doing previously, but it's at a velocity and a scale that was way faster, way higher than before was even possible.
Hey guys, we are covering a ton of ground. And I know you're probably feeling a little overwhelmed and you want to learn even more. Well, good thing we've got a ton of resources for you.
You can scan that QR code or you can click the link in the description below. We'll get all of those for free. You know, Kieran and I wrote this book, Loop, and it's about this exact idea, right?
You're probably running meta campaigns that are performing okay when they start, but, you know, 90 days in are probably performing, you know, twice as good, three times as good as they were when they started, right? Yeah, anecdotally, I'll give you an example. We just are working with a...
private equity company. It's a portfolio that they have. They have 250 physical locations.
We took their cost per lead on Google Ads down from $70 down to 35 in four weeks. Same ad budget. On Facebook, we got their cost per lead down from $80 to $15 is the average seven -day trailing in the same four -week period.
This is just creative volume and basically agents doing negative match keywords on Google Ads and the creative volume on Facebook. This is no different than what a human would do, right? It's just the velocity that that can happen at.
And the scale, it's speed and scale, right? And this is a software engineer running this account, right? It's like basically they're, I almost, I call it like day trading, like ad channels.
They're basically like day trading Facebook ads or day trading like Google ads. It's like how this is happening. It's almost like this like quant, like function that's on top of this technology.
And this is actually like, you know, I think the biggest thing, if you walk away from this conversation today, like I want you to take is that, Your coding agent is now the smartest marketer, the best marketer in the world. If you give it all the context that it needs to do its job.
Maybe we go into that. Maybe pull up an example of one of the paid agents or let's build one live and kind of talk to people. When you say context and data, like what the heck you mean?
And when you say like speed and scale, what do we actually mean? I want to make it real for people who might be still thinking about this world in the older way. So high level, I'll just run through this fast because I know for the sake of time.
But basically, I just want to talk about what is marketing engineering, which we've kind of discussed already, the infrastructure that you need. And then we'll actually go into the deploying of a Google Ads agent, Facebook Ads agent, SEO agent, a cold email agent on top of this. So on the marketing engineering side, we've talked about this already.
All marketing is now code. You should build systems, monitor the systems, find things that work, build the workflow, automate the workflow, audit the workflow, and then improve the system. So what we have found...
Over the last 18 months, what does your coding agent need to actually go and do this marketing engineering or GTM engineering? It's a stack of things. So it's a data pipeline, a data warehouse, a cloud server to host code, media storage, databases for your agents so that they can record their actions that they're doing.
recurring job or tasks based on jobs, application authentication so that you can have your, for example, say if you make a dashboard that you want to have hidden, but shareable to your team, you can do that. The shareable links and then a Git manager. So it can be multiplayer.
So you can have multiple people, say, for example, your five team members working on the same code base at the same time. And then a tailored API gateway that enables you to do calls to Nano Banana or Apollo or Seed Dance or Appify, whatever tooling you need to actually go and do this work. Right.
So the open source version of this, I'll share this with you guys that you can share with the audience. It's basically everything here. But for today, what I'm going to be doing is just using, this is literally what graft is as a platform is everything that I just described here.
And then we're going to go and build some of these agents out, right? So everything I'm going to get access is through the graph CLI, and we're going to build out an agent like this. So what is this agent?
Again, it's some type of software with a thinking loop, with a live data stream that it's making decisions on top of. So the Facebook ads agent basically research the target customer, desired outcomes and pain points. I'm then going to generate creative.
I'm going to automatically upload it to Facebook ads. I'm going to measure the performance. Winners scale, losers get turned off.
The winners influence the next round of creative production. Before we go on, there's a really important thing that I think a lot of people have trouble with, Cody, which is that like generate the creative and auto upload. marketers do not like giving up control.
It's like, wait, no, I made these five things. I know what these five things look like. I want to upload them.
I am not comfortable with like these 50 things that I did not. Well, I guided, I was kind of like the creative director, but I didn't make them. How do you help people get past that way of thinking?
I think what it gets to is like, you're doing this work on the front end. So that you trust the creative that's going to come out the other side, right? Like you're giving your brand style guides, right?
What are our fonts? What are our colors? What's our language that we use?
Like, et cetera. And then also like what we typically do, for example, when we do a forward deployed engineering engagement with a company is we're basically like looking at their historical ads, using that as basically source material. And then also going through like a creative loop with the team to be like, okay, yes, no.
which of this passes, right? You can also build in a gate for a human at this level as well, where it's like every new piece of creative, a Slack message goes out to them. And that person is basically like rejecting or accepting that creative, right?
That is something that's also possible. And then that builds up a repository of like what, you know, the human has okayed, which can then influence the net, like the future creative production. But again, I think it just comes down to the scale.
And like, we see this like, Working with a company where this exact problem is happening right now is like a human was in the loop to begin with. The velocity of creative at this point has gotten so large.
That no human could actually. There's no human. No human can look at a thousand ads a month.
Right. And then also run their other job and give feedback on those thousand ads. Right.
But those ads, when you look at them, like they're on brand, they're talking to the pain points that we've like defined all those pieces. But we're hitting on something really important. There are a lot of people who giving up that control initially is going to be really hard, especially if you're in regulated industry.
There's a whole host of challenges here. I think what we are saying, and that's really important, is one, you can build an infrastructure to get really high quality output. Like, for example, in the Loopbook, we have this concept called a taste profile, Cody, where it's literally all of your customer and brand.
identity, data, insights, emotional triggers, everything in one file that allows an agent like this to make far better decisions. Once you have something like a taste profile, then the thing you do is kind of what you said, is you have a human review and give feedback and annotate for a while. And that a while might be a week, a while might be three months, depends on your company, how conservative or non -conservative you are.
But I think what you found and what I've also found is that in doing all that upfront work, you get pretty close on the creative quality side. It's not going to be perfect.
And there are going to be a few outliers that are imperfect, but the vast majority are going to be very good. Totally. And I think, again, in the regulated space, like human in the loop is a great piece to have.
Yeah, yeah, exactly. Like we work with a company that's in like, you know, the financial markets, right? Like there's just things you can and can't say, right?
Obviously. But you can codify that into the system too. Absolutely.
Absolutely. And like over time, we're like, OK, cool. Here's the language that can and can't be used.
Right. And then also here's the language that like Facebook is more likely to flag when we put it as an ad. So like, let's not use that language.
But this language is like more acceptable. Right. All of that can be built into the system.
And when I say an agent, right, like, again, this is just code that a human can go and modify based off of like their gut intuition, what they're feeling, what they're seeing. Right. We have agents that are just running themselves, right?
Like, for example, we have a Google Ads agent that's running for the startup right now that their cost per lead was $1 ,100 when we started with them. This was four months ago. And we got it down to $250.
What's hilarious is the $1 ,100 was, like, still profitable for them. Like, that would work. Yeah, depending on your model, it can be, right?
Totally. But this agent, basically, what is it doing? It every day is checking the data that's coming back from Google Ads and from Google Analytics.
And then it's rewriting its own software. So it's basically rewriting the software that's deterministic, that's running the Google Ads account. So it's in this infinite loop where it's just basically optimizing itself based off of the live data that's coming from the market as it goes and deploys these ads.
And what is it doing? It's researching the keywords. It's making the ad sets.
It's making the ads. It's making the landing pages that it's sending people to. So it's building that whole funnel optimization basically on top of the back of this.
That was a really important sidebar. Thanks for talking us through with like how you handle this creative and automation at scale. Let's get back to the agents and kind of the core agents that we're walking through here so we can build a couple.
Yeah, absolutely. Yeah. So again, another example of this would be an SEO agent, right?
Same idea, research target keywords that are bottom of funnel related to the brand, research the article that you should write based off of what's ranking on page one. You write the article, publish the article to the CMS via an API, and then you refresh those articles. I think this is the biggest piece that people miss with this type of work is like, Any of these pieces of content, you can go and basically make them better, just like a human should be doing, based off of the live data.
We had a team of people doing this for the last decade. It's super important. Totally, totally.
So with all that said, though, let's actually jump into it, and we can go and do one of these deployments. So I'm going to start with the Facebook ads one. And just for everybody watching, following along on YouTube, you're in Cloud Code here?
Yep. So I'm in Cloud Code. And I've given my agent, my coding agent, Claude Code, basically access to that entire stack that I described previously.
So that data pipeline, that warehouse. So I put all of my data, all of my marketing data into the data warehouse through that data pipeline. Again, I'm using graph .com for this, but you can do this open source as well.
You can use Graph, you can use HubSpot, you can use, there's a ton of different things. Exactly. You got to have those core foundational elements and you have to have a way to get it to what you're calling coding agent, but it's really like you're architected.
They're doing coding. there's system building, all of those things, but they need access to all that. And is that access coming through an MCP server, Cody?
No, so it's through an MPM package that we basically created that gives endpoints to all of... that stack that I described, but also creates this like unified and API endpoint for all the tools that we have within the platform. So as an example, like we've bundled 230 different tools.
What do I mean by tools? Like, for example, if I'm trying to do nano banana creations or I want to scrape a LinkedIn profile or I'm trying to do keyword research for SEO, we're using, for example, under the hood, the data for SEO API for these. All of these endpoints are basically cataloged within that MPM package.
And all I have to do is just say you know with words right like i want to go do this thing let's go and deploy this agent so let's actually do that now i'm just going to use transcription and i'm going to go and say this i'm going to say i want to make some facebook ads for the icp that is basically trying to connect their marketing data to cloud code so that they can do data analysis and for example like a google or facebook ads audit within this so um for the actual creative um use uh go to graph .com and look at the style guide just on the home page of the website and then use the kai api key with a nano banana to create the statics i just want them to be a square format 1080 by 1080.
so i'm going to kick that off Cody, you're doing this in voice, which by the way, just for everybody watching, this is like becoming the de facto way of working. If you're not using voice AI and transcription like that, you should.
And that was like a very casual, like low key way to do this. Like if you were doing this for a client, you'd have different files and different servers you'd be accessing. But I think that was a really great overview of how you would actually prompt to build one of the agent systems you're looking at.
Totally. And I think about it as like co -working with like a team member. Right.
It's basically I'm going back and forth with them. I'm teaching them the thing that I'm trying to do. And then once I've gotten to the point where like, hey, this is actually what I'm looking for.
At that point, I can be like, OK, look at the conversation that we just had. I now want to go and turn this into like a process, into a repeatable process. That is code.
That is an agent. Right. Like what we described previously.
and then deploy that to the server so that that's running on a cron job on a daily task right where it's like daily monthly you know hourly some of them that are functioning in that way so and how i work just to kind of talk through this so this is now working in the background at this point i would kick off another agent and just to show you a different way to do this so we have plugins within graph for example i have an seo plugin all of these plugins people are just starting to publish them to github they're just different agents it's different way to structure agents And again, that agent, what is it?
It's code, thinking loop, live data stream. So a plugin that we have is an SEO plugin. So I'm going to say I want to implement the SEO plugin, look through the graph CLI plugins for that and walk me through the process.
And just for context, I have already given my coding agent access to the Facebook ads API, the Google ads API, my CMS, which we use strappy. It's just like an open source one. And then also to Instantly AI, which is a cold email software.
So that's what's happened beforehand is basically how we can go and do this. And I get this question all the time. So I just want to digress here for a second.
Please, please. I see on social media, Facebook ads accounts getting banned for agents interacting with them. And it's so funny to me because people have this perception that the agent is going to go and that's the reason this is happening.
The reason it's happening is the agent is spamming the API endpoint. And it's breaking the TOS of Facebook ads. And so why is that happening?
It's because it's going and doing like all these reads trying to pull this data in. So how do you solve this? Why do we even need a data pipeline in the data warehouse?
That is the solution for all of the data analysis. So when I am doing data analysis, it's only happening from the data warehouse. And then the only way that I'm really using the API with my coding agent is just for writes.
just for uploading those Facebook ads, just for uploading those keywords or making those new ads within Google ads, et cetera. All right, let's go through what we built. Yeah, so far, again, we're waiting on inference, right?
Which is like a challenge with all of these pieces. But basically in the background, what's happening is Facebook ads is now going and creating the initial media that I described using Nano Banana to actually make the actual images. And Nano Banana is a image generation tool?
Yeah, exactly. By Google and you're paying, you know, per API call for that creative, which is a fraction of what you would pay a human designer to build those. Exactly.
Exactly. And so what it's done is it went to the homepage of the website. It's now scraped out basically what colors that are being used, like descriptions of the actual creative, et cetera.
We're finding that. Honestly, chat GPT image is the best for this. I just use nano banana today because that's what came to mind.
But with this, you can go in these loops, basically making that creative. On the Google ad side, what's happening now is I'm basically having it do bottom of funnel keyword research for the keyword phrases that we should basically be bidding on. It's using the data for SEO API to go and find this.
We also pulled data from our own Google search console because that's connected so that we basically have a list of like the keyword phrases that we can go and We create a campaign that we're going after. Then on the SEO piece, I had it install the plugin.
And at this point, we are capable of publishing an article, right? So for example, I can say we're doing a campaign right now, Facebook ads for X business type. I'll say write an article for Facebook ads for, we'll say pet grooming, publish that article or use the system that's basically defined within the plugin there.
And so what's defined within the plugin is again, that exact thing that I talked about previously, where I'm just giving it a keyword to go and write, but it's going to research this article, write the article, and then actually publish it to Strapi through that API to the live site. So I'm going to do for pet grooming and then publish the article, use Strapi for publishing.
And then on the final piece, just to talk through what we can do on the cold email side. So again, we have cold email that's connected. So we're going to go and basically create a daily cron job that's extracting the net new posts from a creator.
I'm going to give you their LinkedIn profile URL. We're then going to extract the engagers from that post and then do a waterfall enrichment through the email providers that are available through the Graft CLI tools to find the emails, validate the emails with Million Verifier, then upload. those two instantly via the API key.
So again, describing the workflow that I want, and then I'm letting the coding agent use all the tooling that it has available to go and basically build out that workflow. And this is the new way to do this. Like historically, it's like, oh, I would go and try to use N8n or Zapier.
Like the joke that we have or that we keep saying is like, friends don't let friends do N8n. Like go straight to code. That is the best way to do this because it is the most flexible and it's the most capable at connecting all these different data sources.
Why would I go and manually drag and drop a workflow? And then the feedback we always get on this is like, okay, well, I want to visualize it. I need to see the system.
It's like, just ask the coding agent. Be like, visualize the system. Visualize the system for it.
Exactly. And it's going to go and generate basically what that graphic looks like. We demystified agents, right?
Agents, in this case, is just like... A thread in your cloud code, right? You know, we just showed a few different threads, each of those and agents that's doing a task.
And if you do the foundational work of the plumbing, the API access and everything, then you can just describe the system and have that independent agent go and work for you. And I know that can seem intimidating for a lot of people, but with the help of smart technical people and getting the initial setup done, like the ongoing usage is pretty.
pretty manageable, right? You're seeing so many people out there building just stuff because they can build stuff now with agentic coding, right? And the building stuff was never the barrier.
It was, hey, do I really understand the problem to build the right thing? Do I have a group of people who are maniacally focused on what needs to change in the market so that that product changes? Oh, and by the way, am I willing and prepared to do all of the upkeep and maintenance and iteration to actually make this thing good?
And that's the hard part. Like we're working, for example, we just onboarded this company yesterday and it's a marketing team of two. I mean, this is a 15 year old organization, right?
And it's a huge company, but basically like he has told his boss, you know, our contact there has been like, I don't want to hire more staff. I just want to pay for like this tooling basically. That is where the budget is now going to.
And what we're seeing is like this one person. When they have this tooling, you know, that they're surrounded by, it's suddenly 100x employee, right? They can do things that were like, it would take teams of people historically.
And they're also just running circles around everybody that they're competing against. But I'm just curious your guys' thoughts on like what the CMO role, like does that even exist anymore? Is it only like IC?
The summary line for you is essentially the CMO role as it is constructed today is dying because you don't have the right. talent pipeline of people to do what is required and partially is like a lot of cmos for a long time i think we're kind of more professional managers than they were marketers yep and what you're outlining i think is right which is if you're a deep crafts person in marketing you can do the work of 10 or 100 people with enough token spend and the right data access and systems like Those things are non -trivial things, right?
But it is important. And in the book, in Loop, we talk a lot about, like, you want to get rid of context -carrying, people who exist to just move information around. And there's, you know, this idea of, like, a super contributor where, like, a whole person can now own an end -to -end, you know.
meta campaign for the example we're talking about on the show where it used to be well i gotta wait for my freelancer to get me the ads and i gotta wait for my like analyst person to get me the data from the finance team like all that stuff can now happen live and happen with one person and so i think you're seeing that that being said you still need cmos in the future because If I can have one of those 10, 100x people, then I'd have 10 of them or 20 of them or 30 of them, right?
Like there's no bounds to like the upward bounds of growth and you still need a human to manage and orchestrate. Just the ratio of people you need to do the work is changing, I think significantly. It's just like, how many of those people exist today?
Part of the issue is like, that's probably like most optimistic case, sub a thousand. My kind of closeout question to you here, Cody, is like, if I were watching this episode and I've never done any of this, right?
And I'm just like, oh, I'm buying what these two people are selling and that there's a new, better way to do this. Like, where would you start? I think the first thing is just like give, again, your cloud code, your codex, whatever your daily driver is, API keys, like have it stored locally in an environment file and just have it go and try to do work for you.
Like what you would go and manually do, for example, like a... say you're trying to clean up HubSpot or you're trying to do something, get a HubSpot API key and go and try to get your coding agent to do that action for you. That is the first place.
You're going to be shocked by what you're going to be able to get done with that. The other piece I just want to piggyback on top of is like, I just showed basically how to interact with coding agent to like do this initial build and like how I view an agent, right? It's like everything that I just did with it.
You can then take that and again, turn that into the system that's running itself. Yeah. I'm not saying every day you go into cloud code and you're like, okay, cool.
Publish the article. It's like, no, every day, three new articles go live. I deployed that solution into a server and that is just running again on a cron job, running that process that I described.
and then optimizing itself based off of the live data stream that's coming back to it. Cody, this has been a masterclass in how to think about scaling your marketing with agents. You've given us real tactical breakdowns of everything.
Thank you so much for joining us on today's show. I know I got a ton out of it. Before we close out, where can everybody find you if they want to learn more?
Yeah, totally. You can go to graph .com. I'm very active on Twitter and also on LinkedIn, basically documenting everything that we're seeing be functional, also growing a YouTube channel right now.
Thank you so much. We'll see everybody on the next episode. This data is wrong every freaking time.
Have you heard of HubSpot? HubSpot is a CRM platform where everything is fully integrated. Whoa, I can see the client's whole history.
Calls, support tickets, emails, and... Here's a task from three days ago I totally missed. HubSpot.
Grow better.
The Hook
The bait, then the rug-pull.
The episode opens mid-result: a Facebook cost-per-lead that fell from $82 to $15, and a claim that a coding agent, given the right context, is already the best marketer on the team. What follows is a HubSpot podcast interview that turns into a live build, with the guest wiring real ad, SEO, and cold-email agents together in Claude Code by voice.
Frameworks
Named ideas worth stealing.
07:25list
Marketing Engineering Infrastructure
Data pipeline
Data warehouse
Cloud server
Media storage (images, videos)
Databases for agents (Postgres)
Recurring tasks (cron jobs)
Application authentication
Shareable links
Git manager
Tailored API gateway (Nano Banana, Apollo.io, Seed Dance, Apify, etc.)
The ten-part stack Cody says a coding agent needs before it can be trusted to do real marketing work end to end.
Steal forStanding up an internal 'agent ops' stack before wiring any single marketing agent
10:36concept
Taste Profile
A single file combining brand identity, customer data, and emotional triggers so an agent's creative and copy come out on-brand before a human ever reviews it.
Steal forBriefing any generative agent (creative, copy, or landing pages) so its first draft needs less human correction
01:00concept
Agent = code + thinking loop + live data stream
Cody's working definition of an agent: not an autonomous black box, but ordinary software with a decision loop that reacts to a live feed of data, using an LLM only where inference is actually required.
Steal forScoping any 'should this be an agent' conversation down to a concrete build spec
CTA Breakdown
How they asked for the click.
VERBAL ASK
05:20link
“You can scan that QR code or you can click the link in the description below. We'll get all of those for free. You know, Kieran and I wrote this book, Loop.”
Soft, value-first plug for a free resource and the hosts' own book, delivered mid-interview rather than as a hard pitch; the episode also carries a produced HubSpot CRM sponsor spot appended after the sign-off.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A HubSpot marketing lead runs a free Claude Code "marketing team" against a real company's live site and grades every output: a generic headline, a genuinely great one, and the AI audit that faked its own homework.
A 54-minute live demo where Cody Schneider runs seven AI agents simultaneously to build a full GTM machine — ads, outreach, cold email, data analysis — with Greg Isenberg watching.
Ben Senescu built OpenSEO to kill his own Semrush bill, wired it straight into Claude, and now watches open-source contributors fork it into CRMs and content tools he never planned to build.
AI Edge's host installs seven free, high-star GitHub repos live on screen, turning stock Claude Code into a social-research engine, a browser operator, an ads department, and a 279-agent company.
00:17
00:36
01:20
01:44
02:10
02:47
03:22
03:32
03:54
04:22
04:43
04:59
05:19
06:05
06:35
06:42
07:35
07:55
08:15
08:35
09:29
09:35
10:02
10:33
10:50
11:28
11:55
12:36
12:56
13:32
13:43
14:10
14:45
15:02
15:28
15:50
16:11
16:22
16:42
17:36
17:56
18:23
18:39
19:32
19:45
19:56
20:27
20:47
21:08
21:29
21:45
22:32
22:50
23:08
23:28
23:47
24:00
24:53
25:13
25:33
25:57
26:29
26:52
27:26
27:33
21:46
26:03
54:07
19:02
24:49
14:58