How We Built the Growth OS: AI Agents for Marketing

Why we built Growth OS

Every growth marketing agency faces the same tension. Clients want sophisticated analysis and strategic recommendations, but 60-70% of analyst time gets consumed by repetitive data tasks, pulling reports, formatting spreadsheets, identifying anomalies, writing performance summaries. These tasks require expertise to interpret, but not to execute. That makes them the perfect candidate for AI automation.

Growth OS began as an internal project to solve exactly this. The goal was straightforward. Build AI agents that handle the analytical grunt work so our team can focus on the strategic thinking that actually moves the needle for clients.

What is a multi-agent system?

A multi-agent system is an architecture where multiple specialised AI agents collaborate to accomplish complex tasks. Rather than building one monolithic AI that tries to do everything, each agent has a narrow area of expertise and well-defined inputs and outputs. The advantages over a single-agent design are clear. Specialisation, because each agent has a focused prompt and toolset. Reliability, because if one agent fails the others continue. Iterability, because you can improve one agent without touching the rest. And cost efficiency, because smaller, focused agents use fewer tokens and can run on cheaper models for routine work.

What Growth OS covers

Growth OS is organised into specialist teams, each owning one part of growth, coordinated by an orchestrator that runs the queue and keeps them in step. The teams span:

• Search, SEO and AEO: organic and AI-search visibility, built on live Search Console and GA4 data, with technical and content audits and the AEO work aimed at earning citations in AI answers.
• Paid media: Google, Meta and LinkedIn. Budget pacing, bid and creative decisions, and creative-fatigue detection.
• Content: performance analysis, gap-finding, and editorial quality control against the brand's own voice rules.
• Analytics and measurement: cross-platform reporting, anomaly detection, and experiment design.
• Research and ABM: ideal-customer profiling, prospect building, and lead scoring.
• Reporting: client-facing summaries pulled from every source, structured against the client's own KPIs.

The point isn't the number of agents. It's that each one owns a narrow job, proposes the next move, and hands it to a human to approve.

How it is architected

Growth OS runs on a deliberately simple stack. It is the Claude API with tool use, custom MCP servers for scoped access to the ad and analytics platforms, a Next.js application for orchestration and human review, and Supabase for storage and audit logs. One principle sits underneath it all. Agents propose, humans approve. Every recommendation, whether pausing a keyword or adjusting a budget, requires human sign-off before anything happens. AI does the analysis. People make the decisions.

What we learned

Start with the boring tasks. The most valuable AI work automates what is necessary but tedious, not what is creative or strategic. Report compilation, anomaly detection, and data formatting are ideal. Campaign strategy and creative ideation are not, at least not yet.

Prompt engineering is product design. Output quality depends almost entirely on prompt design. We spent more time refining prompts than writing code, and treating them as product specifications, with versioning and testing, was the key insight.

Human-in-the-loop is non-negotiable. We experimented with fully automated actions early on. That was a mistake. The approval step costs 15 minutes of human time and prevents costly errors.

MCP changes everything. Instead of building a custom integration into every agent, you build one MCP server that all agents can use. Adding a new data source means one server, not one integration per agent.

What is next

We are deepening the platform across search, AEO and retention, and tightening the measurement underneath it all. The vision is simple. Give every marketing team the analytical depth of a large agency, powered by agents that work around the clock. The future of marketing operations is not replacing humans with AI. It is amplifying human strategic thinking with AI-powered analysis.