what a marketing agent actually is

most things sold to you as a marketing agent are a zapier workflow wearing a personality. it runs a straight line, top to bottom, and it makes exactly zero decisions.

cody schneider drew the line on greg isenberg's podcast this week, and it's stricter than the marketing around the word. he wanted to separate his thing from the n8n screenshots, so he gave a definition with three requirements, then walked the whole build. the episode ran out of time before he got past the first agent, so the order is conversational rather than constructive.

i read the full transcript and put it back in the order you'd actually build it in.

the short version:

• an agent earns the name when it has unified data, a cadence, and a decision loop. two out of three is automation.
• the definition forces infrastructure on you: a pipeline, a warehouse, and somewhere to host the thing. all four pieces are self-hostable.
• ad accounts get banned for spamming the read side of the API, not for using an agent.
• the failure mode that kills these is entropy. the agent converges on its own best guess and quietly gets worse every day.
• you don't get to pick the winner. the market does, inside 48 hours.

the definition is stricter than you want it to be

three things, and he's firm that you need all three.

first, the data problem is solved: the agent sees the whole pipeline in one place, not one channel at a time. second, it runs on a cadence. third, it makes decisions inside a thinking loop and improves from what comes back.

he draws the contrast in one line. not a linear automation, but "code in the cloud that's making decisions off of your live business data."

what he explicitly does not want is the general-purpose thing that reasons about anything. he wants something narrow that runs a process, reads the result, and adjusts. narrower is the feature, not the compromise.

the definition forces the infrastructure

this is the part people skip, and it's why most attempts stall.

if the agent needs to see the whole pipeline, you need every source in one place. that's a data warehouse. his plain-english version: "a data warehouse is a way to unify all of your data sources into a single location."

to get data into it you need a pipeline. he names airbyte for the pipeline and clickhouse for the warehouse, both open source, both self-hostable.

then the agent has to live somewhere. any cloud host works. his aside here is the practical one: people are running these on a mac mini on their desk and calling it a deployment, and they don't need to be.

for the facebook ads case he lists what has to land in the warehouse before the agent is useful: ad platform data, analytics, product events, crm, and stripe. the stripe row is the one that matters, because without revenue on the same table you can measure clicks but never which specific ad paid for itself.

what the loop actually does

with the warehouse in place, one agent looks like this.

it researches pain points from real complaints, then ranks them by how often they come up. it generates creative against those pain points, statics and video, checked by a vision model against brand rules like fonts and colors. it publishes into the ad account. it lets each batch run two to three days for signal, kills the worst performers, and promotes survivors into a pool where they compete for budget.

then the loop closes. it reads performance back out of the warehouse and uses it to shape the next batch.

the detail i hadn't seen elsewhere: he stores the actual generation prompts and scripts in a database next to the results, so the thing being analyzed is the prompt, not just the ad. that's what lets it get better at making creative rather than only better at picking it.

the api rule that saves your ad account

there's a genre of post about ad accounts getting banned for agent use. his correction is blunt: "the agent is not the reason it got banned."

the bans come from hammering the API to pull hundreds of millions of rows, which is a terms-of-service violation on its own. the fix is a boundary, not a retreat: use the marketing API for writes, publishing and pausing and promoting, and get your reporting from the warehouse you already built.

this is a second, quieter argument for the warehouse. it's not only about giving the agent context. it's what keeps you off the read endpoints that get accounts killed.

entropy is the part that decides if it survives

here's the section i think is the actual episode, and isenberg says on-air that nobody talks about it.

"you have to solve for entropy." the agent doesn't crash. it converges. left alone it keeps producing variations on what already worked, the variations get narrower, and "the agent gets stuck thinking in the same way."

the decay curve is gradual enough to miss. it feels fine on day one, slightly worse on day two, and by the end of the week you have a system confidently repeating itself.

his fixes all amount to pumping novelty in from outside the loop:

1. pull competitor ads out of the platform's public ad library. that "puts new DNA into the system."
2. mine transcripts, youtube and podcasts in your category, for angles your account has never run.
3. scrape what's going viral in short form and ask whether the format transfers.

underneath all three is one idea. a closed loop optimizing against its own history has no source of new ideas, so you have to be the source.

you are not don draper

the last piece is a constraint on you rather than the system.

"we're not don draper." you don't decide what the market wants and impose it. you put many variations in and read which one the market picks, and you know "within 48 hours." his other note here is that most people quit after three or four ads when the number that finds a winner is closer to fifteen or twenty.

and campaigns as a mental model are done. most people still run marketing as "i start a thing and then i stop the thing." the loop doesn't stop.

what i'd change

three things, from running my own loops.

the entropy point is real, and it arrives faster than you'd guess. i built an experiment tracker for my own posting that depended on me tagging each post as i published it. it rotted in four days. not a discipline failure. any loop with a manual step decays to the rate of the manual step, and the fix was not a new system but hanging the refresh off a job that was already running on a schedule. if you take one thing from the episode, take this one, and assume your decay is faster than the version in your head.

i'd think hard before standing up two new pieces of infrastructure. airbyte and clickhouse are both good and both real work to run. if you already have postgres, you can get a long way with it before the warehouse is the thing blocking you. the argument for the warehouse is airtight, but the argument for a specific warehouse is not, and starting with what you run is cheaper than starting correct.

the vision model checking brand compliance is the underrated piece. it's mentioned in passing and it's what makes unattended generation safe to publish. without a check between generation and publishing, you're one bad batch from your brand appearing in something you'd never have approved.

the build order

if you want to try this, the sequence is:

1. pick one channel. one. the episode covers one agent and still runs long.
2. get that channel's data plus revenue into one place you can query.
3. write the loop as a read, a decision, and a write. keep yourself in the middle at first.
4. add the novelty sources before you take yourself out, not after.
5. put a check between generating and publishing.
6. only then let it run on its own cadence.

the full episode is how i use claude code + mcps to run my marketing, cody schneider with greg isenberg. worth the 35 minutes for the warehouse walkthrough alone, which is hard to summarize because he draws it.

start with step two. everything above it is a prerequisite you can't skip, and everything below it is easier than it sounds.