I’m trying to build an AI agency that runs mostly on agents. Not “AI-assisted” — actually agent-run. Here’s the stack I’m testing and what I’ve learned.
The Goal
A business where agents handle client research, content creation, proposal generation, and distribution — while I focus on strategy and final decisions. The benchmark: touch each deliverable once, not ten times.
The Stack
- Claude AI (Claude Code + API) — the brain behind every agent
- n8n — orchestrates every workflow and agent handoff
- Telegram — my command interface. One message triggers everything
- MCP tools — connects Claude directly to Gmail, WordPress, Typefully, and more without custom API code
- Blotato — multi-platform social distribution via API
- WordPress — content hub (manyforce.com)
- Cal.com — client booking and scheduling
How a Client Workflow Actually Runs
A lead books a 30-minute call via Cal.com. Before I get on that call: n8n has triggered a Claude agent, the agent has researched the company and their LinkedIn, a personalised brief has been drafted, and I’ve received a Telegram summary to approve in one tap.
I show up already prepared. No manual research. One approval touch.
How Content Distribution Runs
I send a topic to Telegram. Claude writes platform-specific posts for Twitter, LinkedIn, Instagram, and Facebook. n8n routes each to Blotato via API. Posts go live on schedule.
Total time from me: 90 seconds.
Why MCP Tools Change Everything
MCP (Model Context Protocol) lets Claude agents connect directly to your tools — Gmail, Notion, WordPress, Slack — without writing custom integration code for each one. Without it, every integration is a project. With it, every integration is a config file.
What I’m Still Figuring Out
- Quality control — how much can you trust agents without human review?
- Edge case handling — agents break on unusual inputs
- Client communication — some clients still want a human at key moments
The Honest Take
You don’t need a big team to run an agency anymore. You need the right agents, a solid orchestration layer (n8n), and one human who knows when to step in.
Building this in public. What does your agent stack look like? Drop a comment below.
Leave a Reply