AI can now draft an ad, render a visual, assemble a short video and inspect campaign data in one connected workflow. In Greg Isenberg's conversation with Cody Schneider, the interesting part is not a claim that marketers disappear. It is the environment around the agent: clean data to read, bounded tools to act, creative references to learn from and a person who decides what is worth publishing.
Start with one product and a few draft ads. Give an agent approved brand material and read-only performance data; have it propose angles, produce variants and explain each choice. A human checks claims, likenesses, licensing and visual quality before anything goes live. Only then consider restricted ad-platform write access. The episode demonstrates impressive production speed, but it does not establish conversion lift or a universal $5 cost per finished ad.
Watch Greg Isenberg and Cody Schneider
The video and supplied transcript document what the speakers said and showed. Cody is a co-founder of Graphed, the marketing-agent platform used in the demonstration; treat its product examples and cost claims as vendor claims unless independently measured. Links and repository descriptions checked 10 November 2026.
The Stack Is More Than a Chatbot
Cody's setup separates the source of truth from the action surface. Airbyte moves data from ad platforms, analytics, CRM and payments into a ClickHouse warehouse. The agent can query joined campaign and revenue data there rather than filling its context with repeated ad-API calls. A warehouse does not fix bad attribution, duplicate records or consent problems by itself; those still need a data owner and checks.
| Layer | Job in the episode | Builder note |
|---|---|---|
| Data pipeline and warehouse | Sync Meta ads, analytics, CRM and payment signals; connect creative to downstream revenue. | Check sync freshness, identity matching and attribution before trusting an agent's recommendations. |
| Agent and creative tools | Research a brand, form ad angles, generate images and edit short video. | Give the agent approved brand rules, source assets and a fixed budget for generation and retries. |
| Ad-platform access | Upload, pause or adjust campaigns through platform APIs. | Use least-privilege credentials, a staging or paused campaign, spend caps and a reviewable action log. Platform permissions and rate limits apply. |
| Human approval | Select useful creative, verify claims and decide what runs. | Keep publishing, spend changes and sensitive audiences approval-gated until the workflow is proven. |
Cody describes agents reading from the warehouse while writing to Meta or Google through their respective APIs. That is his architecture, not a recommendation to turn on autonomous budget control on day one. The Meta Marketing API access rules are a reminder that these integrations have real access tiers and limits.
From a Brand URL to Static Ad Drafts
The live exercise starts with brands including Idea Browser, Late Checkout and Cursor. Cody invokes a public static-ads skill: it reads the brand site, reviews examples in Meta's Ad Library, proposes five angles and generates image variants. Greg then critiques whether the output has the brand's voice and whether a creative idea is distinctive enough to stop a scroll. That review is not an optional cosmetic step; the video itself shows why generation and selection are different jobs.
- Define the offer: product, audience, approved benefits, forbidden claims and landing page.
- Collect references: your own winners and licensed visual examples; label why each worked.
- Ask for angles before assets: make the agent write the audience problem, hook and evidence for each concept.
- Generate a small batch: inspect text rendering, brand fit, likeness rights and the first-frame message.
- Review and measure: publish only approved variants, then compare qualified leads or sales as well as clicks.
The repository is useful because its instructions are inspectable. It is not evidence that any of the episode's fresh mockups have won in a live campaign. “Top 0.1% marketer” is a motivating claim in the conversation, not a measured percentile.
The UGC Avatar Demo and Its Real Cost Boundary
The UGC-video skill takes a script through synthetic voice, avatar/video generation, word alignment, timing fixes and motion-graphic finishing. In the episode, the team uses tools including fal, FFmpeg and Mirage Tesseract. Cody checks mismatched spoken words against the script and reruns or repairs sections; that quality pass matters more than the fact that a first draft can render quickly.
The speakers estimate roughly $5 in generation charges for their demo and compare it with hundreds of dollars for a human-produced spot. Those are not equivalent all-in prices. Include API and subscription charges, failed generations, creative direction, manual fixes, usage rights, campaign testing and ad spend before making a buy-versus-build decision. The clip also does not report conversion results.
A synthetic spokesperson must not be presented as a real customer with an invented experience. Obtain permission for any person's likeness or cloned voice, clear music and source footage, and check the rules where the ad will run. For US campaigns, the FTC's testimonial guidance explains why fabricated endorsements and misleading avatar use need particular care. A real human creator may be the better choice when trust and authentic experience are the point of the ad.
Greg's Missing Pieces: Memory and Taste
Greg proposes a swipe warehouse: a searchable store of approved ads, hooks, layouts, audience context and performance notes. It gives an agent better references than a broad prompt like “make it viral.” The store should record source and usage rights, not simply scrape competitors' creatives and invite copying.
His second addition is a creative director agent that scores each concept against a rubric: clarity of the offer, evidence for the claim, distinctiveness, brand fit, legibility, rights and likely landing-page match. This is a proposed review layer in the discussion, not a validated autonomous judge. Use its score to prioritize human review, not to auto-publish. Cody also discusses tighter video quality checks; those are an iteration path, not proof that every rendered frame is already production-ready.
Which Piece Should You Try First?
| If the bottleneck is... | Start with... | Success check |
|---|---|---|
| Slow creative ideation | Three written ad angles from one approved offer, then a small static batch. | A reviewer can identify a distinct audience, hook and supported claim for each. |
| Messy performance reporting | A read-only warehouse query and one reconciled campaign-to-revenue view. | Totals match platform and finance records within an agreed tolerance. |
| Video production time | One consent-cleared UGC draft, not a full campaign launch. | Script, voice, captions, footage and rights pass a frame-by-frame review; total cost is logged. |
| Campaign operations | An agent recommendation and a paused test campaign with human approval. | Every action has a reason, owner, cap and rollback path. |
JQ AI Systems take: a good first experiment is five static drafts for one real offer, one reviewer, one modest paid test and a written postmortem. Record total generation spend, review time, approval rate and qualified conversions. That will tell you more than counting how many images the agent made.
Official Links and Source Notes
Graphed marketing skills repository (inspectable static and UGC workflows) · Graphed platform (vendor product page) · Airbyte documentation · ClickHouse/Airbyte integration · Meta Marketing API access · Mirage Tesseract repository · FTC testimonial guidance. The interview is the source for the live demos and quoted estimates; vendor pages explain tools, not independent campaign outcomes. Checked 10 November 2026.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Introduction |
| 04:29 | Agents as marketers |
| 09:56 | Data pipelines and warehouses |
| 13:56 | Agent access to data |
| 16:15 | Tools for the agent |
| 18:08 | The top 0.1% marketer claim |
| 20:43 | Swipe warehouse and context |
| 22:35 | Creating static ads |
| 29:08 | Static ad results |
| 30:03 | Creating a UGC avatar ad |
| 36:28 | Running the workflow |
| 37:58 | UGC ad results |
| 39:51 | Human creators |
| 41:42 | Reviewing static ads |
| 43:29 | The creative director role |
| 50:49 | Closing thoughts |