← Back

Building a Multi-Agent Marketing System in Singapore

·4 min read·Technical

Why Multi-Agent?

Single-purpose AI tools are everywhere. ChatGPT wrappers with a nice UI. Jasper for copy. Midjourney for images. Each one does its thing in isolation.

The problem? Marketing isn't isolated. Content strategy informs SEO which informs social which informs paid which feeds back into content. It's a system. And if your AI tools don't talk to each other, you're just doing manual work with extra steps.

That's why I built a multi-agent system — not a collection of tools, but a coordinated system where agents share context, pass work between each other, and learn from collective outcomes.

The Architecture

Here's the high-level view:

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│  Research    │────▸│  Strategy    │────▸│  Content     │
│  Agent       │     │  Agent       │     │  Agent       │
└─────────────┘     └──────────────┘     └─────────────┘
       ▴                                        │
       │                                        ▾
┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│  Analytics   │◂────│  Distribution│◂────│  SEO         │
│  Agent       │     │  Agent       │     │  Agent       │
└─────────────┘     └──────────────┘     └─────────────┘

Each agent is specialized but connected. The Analytics Agent's insights feed directly back to Research, creating a continuous improvement loop.

Key Technical Decisions

1. Orchestration Over Autonomy

Early on, I tried letting agents communicate freely — peer-to-peer, each agent deciding when to hand off to the next. It was chaos. Agents would get stuck in loops, duplicate work, or miss handoffs entirely.

The fix: a central orchestrator that manages the workflow. Think of it as the game director — it doesn't do the work, but it decides what happens next based on the current state.

2. Shared Context Store

Every agent reads from and writes to a shared context store. When the Research Agent finds a trending topic, that context is available to every downstream agent. When the Analytics Agent sees a post performing well, that signal propagates back.

This was the single biggest improvement over isolated tools. Context is everything.

3. Human Checkpoints

I built in explicit human review points — moments where the system pauses and asks me to validate, redirect, or approve. These aren't bugs; they're features. The most critical checkpoint is between Strategy and Content, where I validate the editorial direction before any content gets generated.

Lessons From the Asian Market

Building for Singapore and Southeast Asia taught me things you won't find in any Silicon Valley AI playbook:

  • Multilingual isn't optional: My content agents need to handle English, Mandarin, and Malay contexts — not just translation, but cultural framing
  • Platform diversity: WhatsApp, Telegram, LINE, WeChat — the distribution landscape here is fundamentally different from the US's Twitter/LinkedIn focus
  • Trust signals differ: Singapore businesses respond to case studies and government endorsements more than founder stories and hustle culture content
  • Timing is everything: Business communication norms here mean my scheduling agent needed significant recalibration from US-centric defaults

What Failed

Let me be honest about what didn't work:

  • Fully autonomous content publishing: Quality dropped. Hard. I now review everything before it goes live.
  • Generic persona modeling: US-trained models writing for Singapore audiences produced technically correct but culturally flat content. Custom fine-tuning and careful prompting was essential.
  • Over-automating client communication: Clients want to know a human is involved. Transparency about AI usage actually increased trust.

The Results

After 8 months in production:

  • Content output increased 4x with the same team size
  • SEO rankings improved across the board — the Research → SEO feedback loop is powerful
  • Client retention improved because we could deliver more, faster, at higher quality
  • My role shifted from "doing the marketing" to "directing the marketing system"

Build Your Own?

If you're considering building a multi-agent system, here's my advice:

  1. Start with two agents, not six. Get the handoff right between two before adding complexity.
  2. Build human checkpoints from day one. Don't add them later when things break.
  3. Invest in your context store. This is the backbone. Garbage context in = garbage output out.
  4. Test with real work. Not demos. Not toy problems. Real client work, real stakes.

The future of marketing isn't AI or humans. It's AI systems designed to amplify human judgment. And building those systems is the most fun I've had in a decade of web development.

Got questions about multi-agent architectures? I'm always happy to nerd out about this stuff.

← All posts