How to Build an AI Marketing Team That Runs Itself (Live Demo)

Technological concepts & agent framework

Marketing agents as “thinking loops”

Marketing agents are software systems that run a cycle of decisions using a real-time data stream and can take actions in marketing systems.

“Marketing engineering” / GTM engineering as code

Building these systems involves creating end-to-end infrastructure, including:

• Data pipeline + data warehouse
• Cloud-hosted code/server
• Media repository
• Databases for agents to log actions
• Scheduled automation via cron jobs
• Authentication + multi-user collaboration (Git-based workflows)
• A specialized API gateway / unified endpoints to call many marketing tools

Context is the key

An agent becomes effective only when it has the right data + tooling + operational context, for example:

• what creatives look like
• what outcomes are desired
• what constraints exist

Product features / platform approach mentioned

Graph.com / “Claude’s code” workflow (from the live demo)

The described workflow includes:

• Load marketing data into a data warehouse via a pipeline (Graph mentioned; other sources possible).
• Provide coding-agent access through a unified API (described as via an NPM package with endpoints for many tools).
• Allow the agent to generate, publish, and iterate across marketing channels.

Unified multi-tool integration

The platform concept emphasizes combining many tools behind one agent interface, including (examples referenced):

• Nano Banana (image generation)
• Apollo / Seedance / Appify (implied enrichment/campaign tooling)
• SEO API
• Strapi CMS
• Instantly (cold email)
• Ad platform APIs

A demo highlighted combining many tools (e.g., “230”) behind one agent interface.

Key marketing agent use cases (and how they operate)

1. Facebook Ads agent

Creative generation at scale

• Uses an image tool (Nano Banana mentioned; ChatGPT images suggested as potentially strong).

Creative lifecycle / learning cycles

• Creatives stored as JSON blobs (treating static images like data objects).
• Each creative has identifiers (e.g., skew numbers) and is tied to outcomes for agent learning.

Winners/losers loop

• Measure effectiveness
• Scale winners, push losers back
• Winners inform the next creative rounds

Research + pain point targeting

• Agent researches target customers, desired outcomes, and pain points
• Generates creatives aligned to those pain points

Learning-phase caution

• Large launches can get stuck in the “learning phase” unless budget and iteration strategy fit the business.

2. Google Ads agent

Intent-driven decisions

• Uses search term data to determine search intent relevance.

Negative keyword logic

• Based on LLM conclusions:
  • if intent doesn’t match the product, apply negative matches and stop spending on the keyword.

Self-optimizing loop

• Daily checks data (Google Ads + Google Analytics)
• Rewrites/updates deterministic software
• Continues iterating in a loop

3. SEO agent

Bottom-of-funnel keyword focus

• Researches terms that align with later-stage intent.

Publish-and-update cycle

• Identify what ranks on page one
• Write articles
• Publish via CMS API (Strapi referenced)
• Update articles using real data (called out as something many people miss)

4. Cold email agent

Daily scheduled process

• A daily cron job pulls new content (e.g., from a LinkedIn author profile).

Enrichment and verification

• Extracts members from posts
• Runs cascade enrichment
• Finds emails and verifies them (Million Verifier mentioned)

Upload to outbound tool

• Pushes verified contacts to an email tool (Instantly referenced via API key)

Infrastructure requirements & governance / constraints

Data warehouse as the analysis backbone

• Analysis should primarily use warehouse data.
• Agents call platform APIs mainly to write actions (e.g., creating ads), reducing API abuse.

Avoiding platform blocks & compliance

• Misconception addressed: account blocking isn’t because agents “leave,” but because of spamming API endpoints / violating platform terms.
• Mitigation: throttle calls and ensure proper pipeline usage so API requests are limited to necessary actions.

Human-in-the-loop / creative approval

If marketers resist full automation, options include:

• brand style guides (fonts/colors/language)
• using prior ads as source material
• routing creatives to Slack approval/rejection optionally

Regulated industries handling

• Constraints coded into the system (allowed vs. disallowed language/claims).
• “Taste profile” concept: a file capturing customer/brand identity and emotional triggers to improve agent decisions.
• Feedback timing depends on conservativeness—anywhere from 1 week to 3 months.

Performance / review-like claims and metrics mentioned

Facebook Ads example

• Cost per lead reduced from ~$82 to ~$15
• Reported as an average over a 7-day trailing period across a similar timeframe.

Google Ads examples

• Cost per lead reduced from ~$70 to ~$35 over 4 weeks (private equity portfolio).
• Another example: cost per lead from $1,100 down to ~$250 over 4 months (starting point still profitable).

General thesis on what drives results

• Success depends more on creative quantity + iteration speed than on targeting/data alone.
• With advanced AI targeting, heavy targeting inputs may be less necessary.
• Creatives scaled from 5 ads → 50/500+ ads, depending on strategy and learning-cycle design.

Guides / demo workflows emphasized

• Deploy multiple agents on a schedule:
  • start with access/configuration
  • generate initial assets
  • publish via APIs
  • measure
  • run cron-based iterations
  • let the agent optimize with incoming data
• Build workflows without drag-and-drop:
  • contrasted with tools like Zapier/n8n; argued coding-based systems are more flexible for combining data sources
• Use voice/transcription (“Bliss”) for agent instructions:
  • demo used transcription to define tasks like creating Facebook ads for an ICP

Main speakers / sources

• Kodi (Cody) from graph.com: platform/demo speaker; covered agent infrastructure and workflows
• Kieran: author of “Loop”; contributed perspectives and experiences

Referenced tools/sources:

• Graph.com, Claude’s code, Nano Banana, Million Verifier, Instantly, Strapi (CMS)
• Ad/SEO-related APIs (e.g., Facebook Ads API, Google Ads API, Google Search Console, etc.)