Free Technical White Paper

AI Agent Architecture for Marketing Teams

How 45+ Specialized AI Agents Work Together to Run Your Entire Marketing Operation. The Complete Technical Blueprint.

By Joon | AI Agent Architect, Founder of AI Topia · March 2026

45+ Agent Inventory

Architecture Diagrams

Cost & Token Economics

100% Free

No credit card required · Instant PDF access

45+

specialized AI agents

functional categories

13+

integration modules

model cost tiers

Executive Summary

Most companies approach AI marketing by buying a collection of disconnected tools: one for content writing, another for SEO analysis, a third for social media scheduling. Each tool works in isolation. Data doesn't flow between them. There's no shared intelligence.

This white paper documents a fundamentally different approach: a multi-agent architecture where 45+ specialized AI agents collaborate through a shared data layer, each handling a specific marketing function while contributing to a collective intelligence that improves over time.

The system covers the entire marketing operation: keyword research, content planning, writing, review, publishing, performance tracking, competitor monitoring, trend detection, social media content, video scripts, newsletters, and autonomous daily operations. All agents share context, learn from human feedback, and coordinate through orchestration layers.

This paper covers the complete architecture: every agent category, the framework they share, how they communicate, the model tier strategy that keeps costs under control, and the build vs. buy decision framework for technical founders evaluating whether to build this themselves or deploy an existing system.

Why Multi-Agent Architecture

A single monolithic AI can't handle the breadth and depth of marketing operations. Each marketing function requires different context, different tools, and different reasoning patterns. The multi-agent approach mirrors how high-performing marketing teams actually work: specialized roles collaborating through shared systems.

Specialization

Each agent masters one function. A keyword research agent has different tools, prompts, and validation than a content writer.

Composability

Agents chain together into pipelines. Swap one agent for an upgraded version without touching the rest of the system.

Cost Control

Not every task needs GPT-4. Route data collection to cheap models ($0.15/M tokens) and reserve expensive models for creative writing.

Shared Learning

When a human edits a draft, every future agent in the pipeline benefits from that feedback through the preference learning system.

System Overview: 45+ Agents, 8 Categories

SEO Pipeline

Keyword Research, Content Strategy, Writer, Review, Publishing, Performance

Content Creation

LinkedIn, Twitter, Newsletter, Video Script, HeyGen, Image Creator

Orchestration

Auto Mode, Content Director, Opportunity Scoring, Strategist

Intelligence

Performance Collector, Viral Pattern Extractor

Research & Optimization

Content Research, On-Page Optimizer

Assistance

Chat Assistant, Daily Brief Generator

Tool Modules

Supabase, DataForSEO, Firecrawl, WordPress, RAG, Apify, HeyGen

Core Framework

BaseAgent, Settings, Logger

The Agent Framework: BaseAgent Pattern

Every agent inherits from BaseAgent, which provides shared capabilities: preflight checks (validate org, LLM provider, database connection), LLM routing across providers, token tracking and cost attribution, async job management, notification dispatching, and structured workflow logging.

The framework is built on Agno (an open-source LLM agent framework) with FastAPI for the API layer and Supabase (PostgreSQL) as the shared data layer. Agents communicate through the database, not direct calls. This means any agent can be replaced, upgraded, or scaled independently.

Every Agent Implements:

Preflight validation (org, LLM, DB)

Multi-provider LLM routing

Token counting & cost tracking

Background job management

Notification dispatch

Workflow run logging

SEO Pipeline Agents (12)

The SEO pipeline is the backbone. These 12 agents handle the complete lifecycle from keyword discovery through published article monitoring.

Agent	Function	Tier
KeywordResearch	Expand, enrich, filter, cluster, prioritize keywords	Fast
ContentStrategy	Competitor gap analysis + structured article plans	Standard
ContentWriter	Section-by-section drafting with brand voice + RAG	Standard
ContentReview	9-dimension quality scoring against benchmarks	Fast
Publishing	WordPress deploy + media + SEO metadata + KB ingestion	Fast
TrendCollector	10-source signal scanning (Google, Reddit, social, RSS)	Fast
CompetitorMonitor	Sitemap crawling, new article detection, threat scoring	Standard
ContentOpportunity	Multi-signal opportunity scoring + format fit	Standard
TopicCluster	Pillar-cluster-subtopic mapping for topical authority	Standard
Performance	GSC + GA4 synthesis, quick wins, weekly reports	Fast
Optimizer	Content decay detection + optimization actions	Standard
AEO Visibility	AI Overview + featured snippet tracking	Fast

Content Creation Agents (10)

Multi-format content generators. Each retrieves brand voice settings and knowledge base templates via RAG before generating. The ContentDirectorAgent orchestrates which specialist agents to dispatch based on opportunity format-fit scores.

150-300 word posts, hook-body-CTA format

Twitter/X

6-10 tweet threads under 280 chars each

Subject lines + structured section drafts

Video Script

30-90 sec scripts, HeyGen avatar-ready

Article

Long-form SEO articles from plans

Image Creator

Hero + inline images via DALL-E

HeyGen Video

AI avatar video generation + publishing

YouTube Research

Keyword research + competitor analysis

Lead Magnet

PDF generation + landing page copy

Source Materials

Research aggregation for writers

Orchestration Agents (4)

These are the conductors. They don't create content directly. They coordinate other agents, make strategic decisions about what to create and when, and manage the autonomous pipeline.

AutoModeOrchestrator

The nightly pipeline controller. Chains: TrendCollector > CompetitorMonitor > ContentOpportunity > ContentDirector. Applies daily caps, injects learned preferences, logs runs.

ContentDirector

Receives scored opportunities and decides which formats to produce. Dispatches specialist agents in parallel: article, LinkedIn, Twitter, newsletter, video script.

ContentOpportunity

Unified scoring engine. Combines search potential (35%), competitive gap (25%), trend momentum (20%), and engagement (20%) into a single opportunity score with per-format fit.

Strategist

Weekly content strategy: what to create, what to refresh, what to repurpose, what to kill. Based on performance data and opportunity pipeline.

Intelligence & Research Agents (4)

These agents collect, process, and surface intelligence that feeds the orchestration layer. Plus two user-facing agents that provide conversational access to the entire system.

PerformanceCollector

Daily snapshots: keyword positions, traffic, social engagement metrics

ViralPatternExtractor

Identifies viral patterns from high-engagement signals for replication

ChatAssistant

Multi-turn chat interface with access to all data tables, RAG search, and agent dispatch

DailyBrief

Auto-generated morning summary: signals, keywords, content status, social performance

Integration Layer: 13+ Tool Modules

Agents don't call APIs directly. They use shared tool modules that handle authentication, rate limiting, error handling, and response parsing. This means adding a new integration is a single module, not a change to every agent.

Module	Purpose	Used By
Supabase	Database CRUD + vector search	All agents
DataForSEO	Keywords, SERP, Google Trends	SEO pipeline
Google APIs	Search Console + GA4	Performance, Audit
Firecrawl	Web scraping + URL discovery	Competitor, Publishing
WordPress	Post publishing + media	Publishing agent
RAG	Vector search over knowledge base	Writer, Strategy, Chat
Apify	LinkedIn + Twitter scraping	Trend, Content
Perplexity	AI-powered trend summaries	Trend, Research
HeyGen	AI avatar video generation	Video agents
ImageGen	DALL-E image generation	Writer, Image Creator

The Model Tier Strategy

Not every agent needs the most expensive model. The tier strategy routes each agent to the cheapest model that delivers acceptable quality for its specific task.

Fast$0.15-0.80/M tokens(60-80% of all agent runs)

Models: Haiku, GPT-4o-mini, Kimi

Tasks: Data collection, filtering, simple analysis, performance tracking, publishing

Standard$2.50-3.00/M tokens(30-40% of agent runs)

Models: Sonnet, GPT-4o

Tasks: Content writing, planning, competitor analysis, review, orchestration

Advanced$15-75/M tokens(Rare, premium only)

Models: Opus, Sonnet 4

Tasks: Complex strategy synthesis, executive-level reports

Why This Matters for Cost

By routing 60-80% of agent runs to Fast tier models, the system keeps average cost per marketing task under $0.50. A full nightly Auto Mode run (scan + score + draft 5 pieces) typically costs $2-5 in LLM tokens. Compare that to $200+ for a human team doing the same work manually.

Data Flow & Inter-Agent Communication

Agents communicate through the shared Supabase database, not direct API calls. This decoupled architecture means agents can run independently, at different times, on different servers, and still collaborate through shared state.

Key Data Tables

keywords

Research data, priority scores, content status

article_plans

Structured outlines, sections, FAQs

articles

Drafts, published content, WordPress IDs

content_drafts

Multi-format drafts with review status

content_opportunities

Scored opportunities with format fit

trend_signals

Signals from 10+ sources

client_preferences

Learned preferences from feedback

workflow_runs

Token usage, cost, duration per run

Cost Architecture & Token Economics

Every agent run logs input tokens, output tokens, estimated cost, and duration. This creates full visibility into where money is being spent and which agents deliver the best ROI.

Operation	Avg Cost	Manual Equivalent
Full keyword research batch	$0.30-0.80	$200+ (analyst hours)
Article plan with gap analysis	$0.50-1.00	$150+ (strategist)
Full article draft (2000+ words)	$1.00-2.50	$300+ (writer)
Content review & scoring	$0.20-0.50	$100+ (editor)
Nightly Auto Mode run	$2.00-5.00	$500+ (team day)
Weekly performance report	$0.30-0.70	$200+ (analyst)

Build vs. Buy Decision Framework

If you're a technical founder evaluating whether to build this yourself, here's the honest breakdown:

Build If:

+You have 3-6 months of engineering time
+You need deep customization of agent logic
+You want full control over LLM provider costs
+Your marketing stack is highly unique
+You have in-house AI/ML engineering talent

Buy If:

+You need results in weeks, not months
+Your team is marketing-first, not engineering-first
+You want proven scoring formulas and pipelines
+You'd rather focus on strategy than infrastructure
+You need the feedback loop and preference learning

Get the Complete Architecture Guide

Download the full white paper with detailed agent specifications, complete data flow diagrams, database schemas, cost breakdowns, and the implementation guide.