Hey Sonny here! 👋🏼
I've been meaning to write about this for a few weeks now but I wanted to make sure it actually worked first.
I built a team of 8 AI agents that run different parts of my business. Not one chatbot trying to do everything — I tried that for months and it was chaos. Eight specialized agents, each with a specific job, working together.
Let me break it down 👇🏼
So here's what I built 🤖
You know how a startup doesn't hire one person to do engineering, marketing, QA, ops, and research? Same logic. One "do everything" AI assistant falls apart at scale. The breakthrough was specialization.
I gave each agent one job — a copywriter for scripts and captions, a developer that picks up Linear tickets and creates PRs, a QA agent that reviews everything before it goes live, a research agent for deep dives, an ops agent for sponsors and deadlines, and an automation agent for pipelines.
The whole thing runs on OpenClaw — an open-source framework for running AI agents locally. I have one main agent called Neo that acts as the coordinator. The rest are sub-agents that Neo spins up on demand. When I send a message, Neo decides which specialist to delegate to, spawns them with full context, and they do the work in their own isolated session.
So when I say "write today's newsletter," Neo spawns the copywriter with the topic and context. When that's done, Neo spawns the QA agent to review it against a checklist. Each sub-agent has its own workspace, memory, and model config — connected to GitHub, Linear, Notion, and Clerk via MCP servers.
The key insight: the main agent IS the coordinator. It's not a separate routing layer — it's the agent I talk to every day, and it knows when to handle something itself vs delegate to a specialist. That's what makes it feel seamless rather than managing 8 separate chatbots.
I just released a video yesterday showing how to set up OpenClaw securely on a VPS so it runs 24/7. If you want to build something like this yourself, that's where I'd start.
Watch this to set things up👇🏼
Watch my latest YouTube video to learn how to set one up in 20 minutes, securely!
Use this link when you sign up for the VPS to get 15% off!
|
Steps to get started!
- Set up OpenClaw on a VPS — one agent running 24/7 is already a game-changer. I walk through the full secure setup in the video linked below.
- Connect it to your tools via MCP servers. GitHub, Linear, Notion, Clerk — whatever you use daily. Once your AI can actually DO things, not just chat, everything changes.
- Make it composable. Skills work best when they do ONE thing really well. Build small, focused helpers your agent can chain together. Check out skills.sh and Clawhub (dont forget to ask your agent to scan for malicious / prompt injection attacks before installing!)
When NOT to use this: Don't hand off things that need human judgment — architecture decisions, design taste. Agents are great at execution, not judgment calls.
What else is going on in Tech?
Wild week. Quick hits:
- 🔮 Claude Opus 4.6 dropped — adaptive thinking, 1M context window, highest agentic coding scores ever. It's what powers my entire agent team. GitHub Copilot already has a 2.5x fast mode for it.
- 🤖 GitHub Copilot shipped GPT-5.3-Codex — 25% faster on agentic tasks. Claude AND GPT competing inside the same IDE now.
- 💻 NVIDIA: 30,000 engineers using Cursor daily. 3x code output, flat bug rates. That's not a pilot — that's company-wide.
- 🤯 Anthropic's CPO: nearly 100% of their code is now AI-generated. The company that builds Claude writes all their code with Claude.
- 📉 ChatGPT market share: 69% → 45% in one year. Gemini surged to 25%.
- ⚡ TypeScript 7 coming — full rewrite in Go, up to 10x faster builds. Already previewing in VS Code.
Are you using AI in your dev workflow yet?
Hit reply — what's working or broken?
I read every single response.
Talk soon PAPAFAM,
Sonny 👋🏼
Come say hi & follow on socials:
TikTok → tiktok.com/@sonny.sangha
Responses