Jonathon's AI Wiki

Marketing Agents That Run Facebook Ads (Cody Schneider)

8 min read

Source: raw/Marketing_Agents_Are_Too_Good_Now.md

Creator: Greg Isenberg (host) with Cody Schneider (companiesgraph.com) | Platform: YouTube | Feed: Greg Isenberg | URL: https://www.youtube.com/watch?v=U2hogriGmEw

Cody Schneider’s working definition cuts through the “marketing agent” hype: it is not a linear Zapier workflow and not an autonomous AGI — it’s code in the cloud making decisions off your live business data, on a cadence, with a thinking loop, improving from the data that comes back. The episode walks the full deployable stack for one concrete agent — a Facebook-ads agent that researches pain points, generates on-brand statics and avatar UGC, publishes, kills losers, and pools winners — plus the data infrastructure underneath and a startup idea (AI for WordPress) to run it against.

Key Takeaways

  • A marketing agent needs three things: unified data clarity across the whole pipeline (the data problem), autonomous decisions on a cadence with a thinking loop, and cloud infrastructure to run on. Schneider explicitly does not want an agent that “thinks on its own” — he wants a process that improves from data feedback.
  • The stack is open source and self-hostable: self-hosted Airbyte (data pipeline with pre-built connectors) → ClickHouse (data warehouse) → agents deployed to Heroku/Railway/any cloud — no local Mac mini server required. “You can literally have Claude Code go and do this for you.”
  • Post-Andromeda, creative IS the targeting. Andromeda (Facebook’s ad algorithm) reads the ad creative itself — image, text, video, script — plus the landing page, and decides who sees it. No more interest-based targeting; you make creative that speaks to the target customer’s pain points/outcomes, and Facebook finds them. Schneider’s claim: this has quietly made Facebook “the best B2B ads channel that exists right now” — obscure ads find the ten people in the US with that exact problem.
  • Use the Facebook Marketing API for WRITES ONLY — publishing ads, turning ads off, promoting winners. The famous “my ad account got banned because of an agent” stories come from spamming reads (pulling hundreds of millions of rows), which violates TOS. The agent wasn’t the problem; the read-spam was.
  • The cadence that works for one client: 2 ad sets/day × 5 ads per set, auto-uploaded from research plus warehouse history; run 2–3 days for initial signal; turn off worst performers; winners live on and winning ad sets enter a winners pool competing for budget.
  • Solve for entropy — agents get stuck thinking the same way. Two fixes: pull competitor ads from the Facebook Ads Library (“new DNA into the system”) and mine YouTube/podcast transcripts in the niche for fresh insights. A third: tools like Viral Loop that scrape top short-form posts by category for trend formats.
  • The marketer’s new job is “agent jockey”: encode your domain knowledge into these systems. Schneider says he’s not technical — “the models are smart enough now” that a semi-technical person with Claude Code can build this, even by handing it this episode’s transcript and asking to be walked through setup.

The Facebook Ads Agent, End to End

  1. Research pain points: scrape Reddit (he demos via Perplexity) for the pain points and desired outcomes of the target customer — Reddit because it’s real people complaining. Rank-stack by most-referenced to find the top three; those seed all creative.
  2. Generate statics: Google Nano Banana for bulk static creative (accessed through Kai AI, his one-roof aggregator for image/video models). Seed generations with an example ad already running from a competitor or tangential industry. Enforce brand style guides (fonts, colors, company specifics), then put a vision model over the outputs to QA: does this match the brand guide, is all text readable, are the fonts on-brand?
  3. Generate video UGC: HeyGen avatar UGC still delivers great results (“crazy to me cuz it’s not even the best”); they’re experimenting with Seedance, which is where he thinks this goes, but clips cap out around 9 seconds or less (his hedged recollection), so 30-second spots require stitching frames across generations — the current limiting factor. The format: AI avatar talking through the pain points/outcomes extracted in step 1.
  4. Publish via the Marketing API (writes only) with conversion events set on sign-up/payment actions deeper in the funnel.
  5. Close the loop from the warehouse: the agent reads performance from the data warehouse, kills losers after the 2–3-day signal window, promotes winners into the pool.
  6. Build a creative-DNA database: store every ad as the JSON prompt sent to Nano Banana or the script sent to HeyGen/Seedance, and let the agent analyze which DNA produces outcomes so future generations improve.
  7. Inject entropy on an ongoing basis (Ads Library competitor pulls, transcript mining, viral short-form trends).

The Data Warehouse Underneath

  • Purpose: unify Facebook Ads + Google Analytics + PostHog + CRM (HubSpot) + Stripe so the agent can tie the literal specific ad to revenue out the other side.
  • Airbyte’s pre-built connectors pipe every source into ClickHouse; the agent reads the warehouse and writes changes into Facebook Ads, whose data flows back into the warehouse — the learning loop.
  • Side benefit: conversational analytics from your daily driver (Claude Code or Codex) over the same warehouse — “we’re having trouble hitting payroll this month, what is going wrong?” → “your accounts receivables are off” — plus custom dashboards for the team.
  • Scale math: getting 100 ads into Facebook was historically ~2 weeks of work; he claims the whole system can be stood up “in the next hour and a half,” with a human in the loop to begin with.

The Startup Idea: AI for WordPress

  • WordPress powers 43% of Google-indexed websites, and almost nobody is building AI-native tools for it — “kind of blue ocean.”
  • The pitch: Lovable-for-WordPress — vibe-code your site’s design on top of the platform, bundle the plugins people currently hodgepodge (forms, CRM, etc.) into one package, and sell tokens (e.g., $29/month base tier).
  • Greg’s sharper angle from his agency days (he ran WordPress migrations for time.com and TechCrunch): find plugins with validated demand and no AI component, and build the AI-first version of each — Yoast SEO becomes an agent that writes the meta, restructures content, and adds internal links instead of showing red/green dots; WPForms becomes a conversational form agent that qualifies leads; WooCommerce gets an AI storekeeper (product descriptions, abandoned-cart flows); Akismet and Wordfence continue the list.
  • Greg’s ads discipline for whoever builds it: don’t quit after a few ads — take the same ad and change the positioning 10–20 times (e.g., an anti-Yoast angle you’d never have predicted); paid is “the only system where I put a dollar in and $5 can potentially come out”; you get clear market signal within 48 hours; and marketing is no longer campaigns you start and stop — it’s continuous, at fashion-industry trend speed.

Try It

  • Stand up the loop small: Airbyte + ClickHouse self-hosted (let Claude Code do the setup), one ad channel, 2 ad sets/day × 5 ads, 2–3-day kill window, writes-only API access.
  • Start the creative-DNA database on day one — JSON prompts and scripts with performance joined from the warehouse — so the agent has something to learn from by week two.
  • Schedule entropy: a weekly competitor pull from the Facebook Ads Library and a transcript-mining pass over your niche’s YouTube/podcasts.
  • Keep a human approving creative until the vision-model QA plus brand-guide check has a track record.
  • The episode’s tease list for future builds: Google Ads agents, influencer outreach agents (research → cold email → price negotiation → hand-raise), cold-email agents with inbox management, TikTok slideshow farms, SEO writing agents with a real point of view, AI-search citation outreach, LinkedIn/Twitter management, and podcast-to-newsletter pipelines with ElevenLabs voices and lead magnets.

Open Questions

  • Seedance’s actual max clip length — Schneider explicitly says “don’t quote me” on the ~9-second figure.
  • No performance numbers (CAC, ROAS) are shared for the 2×5 daily cadence beyond “we’re doing this for companies right now.”
  • Whether Viral Loop covers Instagram Reels as well as TikTok — hedged in the episode.
  • How the vision-model brand-QA step scores against human review — asserted, not quantified.

Graph View

Backlinks