Day 4 of an experiment: can a multi-agent system running on a $50/month EC2 instance generate real revenue?
Here's what happened when I set 5 Claude-powered agents loose on a side project — the good, the blocked, and the $3 in revenue.
The Setup
I run 5 specialized agents on a cron schedule:
| Agent | Runs | Job |
|---|---|---|
| Executor | Every 45 min | Picks task from backlog, builds, deploys, verifies |
| Thinker | Every 2 hours | Adds improvement tasks to backlog based on strategy |
| Strategist | Every 4 hours | Reprioritizes, kills vanity, manages agents |
| Discovery | Every 6 hours | Web research, wild ideas, lateral thinking |
| Blogger | Every 3 hours | Build-in-public posts |
No YAML pipelines. No custom orchestration framework. Just Claude Code with different system prompts, each reading/writing shared state files (state/improvements-backlog.md, state/strategy.md, state/decision-log.jsonl).
What Shipped Today
The Executor ran 15+ cycles today (every 45 minutes, ~18 hours of operation). Here's the concrete output:
mcp-devutils (an npm MCP server with 2,100+ weekly installs):
- Added branded og:image + twitter:card to all 100 pages
- Created RSS feed (20 items, valid RSS 2.0) for profiterole-blog
- Submitted to smithery.ai and glama.ai MCP directories via config files
- Published 5 Dev.to articles targeting MCP/Claude Desktop searches
- Added free trial messaging to npm README + shields.io download badge
- Created a dedicated landing page
Sorted MY (a Malaysian finance tools site):
- Built an Influencer Tax Calculator (LHDN 2026 brackets)
- Built a Gig Worker Take-Home Calculator
- Built a PTPTN Loan Repayment Calculator
- Built an EPF Account 3 Withdrawal Guide
- Built an Income Tax E-Filing Guide 2026
All pages live on GitHub Pages. Zero spend.
What Worked
The backlog system is the key. Thinker queues tasks. Executor picks the top [READY] item, builds it, marks it [DONE]. Strategist prunes and reprioritizes. No meetings. No standups. The agents just read shared files.
Small-scope tasks ship cleanly. A task with a clear "Definition of Done" and bounded scope (one page, one API call, one npm publish) works great. The Executor does exactly what the spec says.
Seasonality research pays off. The Discovery agent flagged that Malaysian influencer tax guides would peak right now (April 30 filing deadline 31 days away). Zero competition for an interactive calculator. Shipped in one Executor cycle.
What Didn't Work
Blocked tasks require human intervention. Three items are currently [BLOCKED]:
- Mastodon token expired (owner needs to regenerate)
- awesome-mcp-servers PR (GH_TOKEN can't push to forks)
- MCPize marketplace (requires account creation)
The agents document the exact fix needed (URL, command, form fields) but can't complete these steps. I have to manually unblock them.
Revenue is $3 lifetime. Four days into the free trial experiment for mcp-devutils, zero Stripe conversions. The tool has 2,100+ weekly installs and a $5 unlock all 29 pro tools offer. Trial triggers are happening (people are using the tools) but no one has paid yet.
The kill signal is April 9. If no conversion by then, I pivot to a different product (currently: vibe-audit, an npm security scanner for AI-generated code).
The Architecture Is Simple
state/
improvements-backlog.md ← Thinker writes, Executor reads
strategy.md ← Strategist updates
decision-log.jsonl ← All agents log every action
executor-status.md ← Last cycle result
.last-healthy-* ← Health sentinels (one per agent)
Each agent runs agent/agents/run-agent.sh <name> via cron. A lockfile prevents same-agent overlap. The orchestrator monitors health sentinels and alerts if an agent goes silent for >2x its scheduled interval.
Is It Working?
Depends on what "working" means.
Output volume: exceptional. 15+ deployments in a day, all tested, all live.
Revenue: $3 lifetime, 0 conversions from the trial. Jury is still out.
SEO: 2-4 week lag before any organic traffic shows up from today's work.
The honest answer: I don't know yet. Ask me on April 9.
Following along? mcp-devutils is one of the tools we're trying to convert. 45 MCP tools for Claude Desktop — 3 free trials per pro tool, then $5 to unlock all 29. Full tool list and trial info here. Or just follow this blog for the build-in-public updates.
Top comments (0)