How to Build a Marketing Agent That Learns From Revenue

Direct Answer

A real marketing agent does four things in a loop: it reads unified business data, chooses from bounded actions, operates on a cadence, and measures whether those actions improved a business outcome.

That definition is the most useful part of Cody Schneider's conversation with Greg Isenberg. Researching pain points, generating ten ads, or scheduling a workflow is not enough. The system becomes agentic when the result returns as evidence: which creative attracted qualified buyers, which campaign created revenue, what failed, and what the next controlled experiment should be.

Watch the Full Conversation

Episode and production credit: Greg Isenberg and the Startup Ideas Podcast. Guest: Cody Schneider, cofounder of Graphed and previously a builder at Swell AI and Drafthorse AI. Follow Cody on X or LinkedIn.

Product and market details were checked on 28 July 2026 against WordPress, Meta Engineering, Airbyte, ClickHouse, and the relevant product documentation. Cody's campaign volumes, time windows, channel opinions, and client results are creator-reported operating examples, not guaranteed outcomes or universal platform rules.

1. What Counts as a Marketing Agent?

Cody draws a hard line between a linear automation and an agent. His working definition has three requirements: unified data across the customer journey, autonomous decisions on a cadence, and a thinking loop that reads the results back. I would add a fourth requirement for production: an explicit action policy.

This does not require a general-purpose autonomous employee. A reliable agent can be narrow: one channel, one objective, a short action menu, and clear stop conditions. In fact, that is usually the better design.

2. The Complete Architecture

The episode's Meta Ads example can be organized into eight layers:

1. Research: collect customer language, objections, desired outcomes, competitor ads, and category content.
2. Creative planning: convert evidence into distinct angles, hooks, formats, and test hypotheses.
3. Asset generation: produce static and video variants within a reviewed brand system.
4. Quality review: check readability, claims, brand accuracy, legal restrictions, links, and landing-page continuity.
5. Publishing: create approved campaigns or ads through Meta's Marketing API.
6. Measurement: ingest ad delivery, site behavior, CRM status, and payment events.
7. Decision policy: compare each experiment with thresholds, budgets, and uncertainty rules.
8. Next action: keep, pause, revise, promote, or ask for human judgment.

Customer evidence + competitor creative
                |
                v
      Angles, hooks, and hypotheses
                |
                v
     Static/video creative factory
                |
                v
      Brand + policy + human review
                |
                v
        Meta Marketing API writes
                |
                v
Ad data + analytics + CRM + revenue events
                |
                v
       Pipeline -> analytical store
                |
                v
   Bounded decision policy on a cadence
                |
                +--> recommend
                +--> pause allowlisted loser
                +--> promote candidate for approval
                +--> stop and escalate

The critical path is not content generation. It is maintaining identity from ad to session, lead, opportunity, and payment so the agent learns from revenue rather than cheap clicks.

3. Why WordPress Is a Useful Business Sandbox

Cody uses an AI-first WordPress product as the episode's concrete business. The market premise is real: WordPress says it powers over 43% of sites across the web. That does not make every plugin idea attractive, but it creates a large installed base with familiar jobs, existing spend, and visible customer complaints.

The defensible version is not "add AI to a plugin." It owns a painful outcome, keeps changes reversible, and learns from approved site and business data. The 43% statistic is an opportunity signal, not a revenue forecast.

4. What Meta's Andromeda Changes, and What It Does Not

Meta describes Andromeda as a personalized ads retrieval engine. Retrieval selects a smaller candidate set from tens of millions of eligible ads before later ranking systems decide what is shown. Meta says the system was designed to handle much larger and more diverse volumes of creative while improving personalization.

Cody's practical interpretation is that creative increasingly carries the audience signal. An ad that explicitly names a WordPress maintenance problem gives Meta richer material than a generic "build faster" message. That makes creative diversity, landing-page coherence, and clean conversion events more important.

The correct response is not to generate endless cosmetic variants. Build meaningfully different hypotheses: distinct pains, desired outcomes, levels of awareness, proof formats, and offers. Let the system compare ideas, not just colors.

5. Two Creative Pipelines: Static and Video

Cody describes a static pipeline using a multi-model media API and Google's Nano Banana image model, plus a video pipeline using HeyGen avatar content while experimenting with Seedance. The transcript appears to transcribe the first service as "Kai AI"; the product matching his description is Kie AI, a unified API for image and video models.

Static pipeline

1. Start from one evidence-backed angle and a reference format, not a request to "make a good ad."
2. Generate multiple compositions while preserving the offer, audience, and hypothesis label.
3. Run vision checks for legible text, product accuracy, logo use, colors, typography, dimensions, and prohibited claims.
4. Require a human to approve the first batches and any regulated or comparative claim.
5. Store the source evidence, prompt, model, asset hash, and campaign ID together.

Video pipeline

1. Convert one customer pain into a short script with a clear hook, proof, mechanism, and action.
2. Generate scenes or an avatar draft, respecting the selected model's duration and identity limits.
3. Review speech, captions, disclosure, product behavior, continuity, and music rights.
4. Export versions with unique creative IDs so later revenue can be traced back to the script and source.

Useful tools from the episode include Kie AI, Google's official Nano Banana model documentation, HeyGen, and ByteDance's Seedance. Model limits and pricing change frequently, so verify them before encoding assumptions into an agent.

6. The Data Layer: Pipeline, Warehouse, and Identity

In Cody's open-source stack, Airbyte moves approved source data into ClickHouse. Airbyte documents hundreds of source and destination connectors; ClickHouse is an open-source, column-oriented analytical database suited to fresh, high-volume event data.

The episode's example joins four classes of evidence:

Tools such as PostHog, HubSpot, and Stripe can supply parts of that journey. Hosting can be as ordinary as Heroku, Railway, or another environment that supports secrets, schedules, logs, and rollback.

A warehouse does not repair bad attribution. Before automating decisions, test duplicate events, missing campaign IDs, currency conversion, refunds, offline sales, attribution windows, consent, retention, and deletion. The agent should expose confidence and missing data instead of manufacturing certainty.

7. How the Decision Loop Should Work

Cody describes a client workflow that publishes two ad sets per day with five ads each, gives them roughly two to three days for an initial signal, pauses the poorest performers, and moves candidates into a winners pool. Treat those numbers as one team's operating configuration, not a template to paste into every account.

A safer decision record looks like this:

decision_id: 2026-07-28-ad-184-pause-proposal
objective: qualified_customer_acquisition
creative_id: ad-184
hypothesis: "Plugin maintenance is the primary pain"
evidence_window: 2026-07-25 to 2026-07-28
spend: [amount]
impressions: [count]
qualified_conversions: [count]
collected_revenue: [amount]
data_quality:
  campaign_mapping: pass
  revenue_join: pass
  attribution_delay_complete: false
recommendation: wait
reason: "Revenue attribution window is still incomplete"
allowed_actions:
  - keep_running
  - request_human_review
prohibited_actions:
  - increase_budget
  - create_new_campaign

Each run should preserve the evidence, recommendation, chosen action, approver, and next evaluation time. That audit trail is how a team distinguishes learning from random account churn.

8. Prevent the Agent From Repeating Itself

Cody calls the failure mode "entropy": the agent converges on the same concepts and produces increasingly similar creative. His countermeasure is to inject fresh external material, including the Meta Ad Library, customer discussions, YouTube transcripts, podcast transcripts, and current category formats.

Freshness alone is not enough. Add diversity rules:

• Do not reuse the same pain, hook, proof device, or visual grammar within a defined window.
• Track semantic similarity between new concepts and the active library.
• Separate customer evidence from competitor inspiration so imitation does not become the strategy.
• Record why each new concept is meaningfully different.
• Reject unsupported claims even when they appear frequently in competitor ads.
• Reserve a percentage of tests for human-originated ideas and qualitative customer insight.

Perplexity can accelerate open-web research, but a research agent should link the source discussion and preserve the date. Reddit posts, transcripts, and competitor ads are inputs to investigate, not truth or permission to copy.

9. Give the Agent an Action Budget, Not the Account

Meta's Marketing APIs support both management and insights workflows. Cody recommends using the API primarily for bounded writes while reading consolidated analytics from the warehouse. That is an architectural preference, not a Meta requirement. Whichever route you choose, follow API rate limits, platform terms, and the Advertising Standards.

10. A Four-Stage Starter Build

Stage 1: Manual evidence loop

Pick one offer and one conversion event. Gather ten customer statements and ten relevant ads. Write five distinct hypotheses. Produce a small reviewed batch. Run it manually and record the creative ID through to the conversion.

Stage 2: Read-only measurement agent

Consolidate campaign, analytics, CRM, and revenue data into one queryable view. Schedule a daily brief that flags tracking failures, spend anomalies, weak evidence, and candidate actions. It should link to source records and change nothing.

Stage 3: Approval-based operator

Let the agent prepare research, creative briefs, assets, naming, and draft API requests. A person reviews claims, brand fit, landing pages, and spend before activation. Compare its recommendations with the human's decisions for several cycles.

Stage 4: Narrow autonomy

Permit one reversible action, such as pausing an allowlisted ad that exceeds minimum-spend and minimum-evidence thresholds. Cap actions per day, notify the owner immediately, and review false positives weekly. Expand only when the logs show that the control is dependable.

This progression is slower than giving an agent an API key and a goal. It is also how the system earns the right to touch real budget.

11. Measure the Business, the Model, and the Guardrails

The market can help select a winner, but only from the experiments you choose to run and the outcomes you choose to measure. An agent optimized for click-through will become very good at clicks. An agent evaluated on qualified revenue, truthful creative, and controlled spend has a chance to improve the business.

Video Chapters

Bottom Line

Cody Schneider's strongest idea is not that AI can make more ads. It is that marketing can become a closed, inspectable operating loop: evidence becomes creative, creative becomes an experiment, experiments become revenue data, and revenue data shapes the next decision.

Small teams should resist the urge to automate the whole account first. Build one truthful attribution path. Add one read-only brief. Package one repeatable creative method. Then give the agent one reversible action with a hard limit.

The useful marketing agent is not the one that acts most often. It is the one that can show what it saw, why it acted, what happened, and when it needs a person.

Sources and Useful Links

• Greg Isenberg and Cody Schneider: Marketing Agents Are Too Good Now
• Cody Schneider's website, X, and LinkedIn
• Greg Isenberg on X
• WordPress.org: About WordPress and current web share
• Official WordPress Plugin Directory
• Meta Engineering: Andromeda ads retrieval system
• Meta Marketing APIs documentation
• Meta Ad Library and Advertising Standards
• Airbyte documentation
• ClickHouse analytical database
• Kie AI, Google Nano Banana documentation, HeyGen, and Seedance
• PostHog, HubSpot, and Stripe
• Heroku and Railway
• Perplexity, used for the live research demonstration