On April 16, 2026, I started building the marketing system. By April 29, it had 659 commits and was running autonomously.
14+ AI agents. An event bus for coordination. Cron scheduling. A prospect pipeline backed by SQLite. An email engine with warm-up scheduling and approval gates. A content generation system feeding 5 Hugo websites. A social media engine posting to X and LinkedIn in my voice 3-4 times a day. A dashboard with 10+ apps. Health monitoring. Performance analytics. Daily briefings.
One person built this. Here’s how.
The stack
Everything runs on Docker containers on a VPS. The agents are Claude Code sessions triggered by cron jobs. They read from and write to a shared SQLite database. An event bus keeps them coordinated so they don’t step on each other.
The websites are Hugo static sites deployed to Cloudflare Pages. The email engine uses Zoho SMTP for cold outreach and Postmark for opted-in subscribers. The social engine hits the X API and LinkedIn API directly.
No frameworks. No Langchain. No vector databases for the marketing side. Just SQLite, shell scripts, Node.js, and a lot of prompt engineering.
The agents
Each agent has a specific job:
- Prospect Researcher. Finds and qualifies prospects matching our ICP profiles using web search and Apollo.
- Email Engine. Composes personalized outreach, manages warm-up sequences (5/day ramping to 50/day), handles replies using NEPQ methodology.
- Content Strategist. Identifies topics worth writing about, generates blog drafts for review. Topic clusters, keyword gaps, regulatory changes.
- Social Engine. Posts to X and LinkedIn 3-4 times daily in my voice. Weekly long-form articles. Promotes new blog content automatically.
- Performance Evaluator. Measures everything. Open rates, click rates, reply rates across ICP types, sequence steps, subject line patterns.
- Pipeline Health. Runs 23 checks every 6 hours. Fixes safe issues automatically, escalates the rest.
- Daily Briefing. Compiles overnight activity into a morning summary.
- Strategic Orchestrator. Reads the event stream for patterns. Spots what’s working and what’s breaking.
Plus a lead orchestrator, work logger, and agent router for @mention dispatch.
The approval gate
Nothing sends without my review. The agents compose drafts. I approve or reject through the dashboard. Only approved content goes out. This isn’t a fully autonomous marketing operation. It’s an autonomous draft-and-prepare operation with a human gate on everything that touches the outside world.
659 commits
That number is real. AI coding tools make this velocity possible for someone who isn’t a career developer. Each commit is small, focused, tested. Describe what you want, build it, test it, commit, move on.
The system is pre-revenue. $0 from marketing tools so far. I’m posting that number because building in public means showing the parts that aren’t working yet too.