Running a small tech services company, I faced the classic scaling problem: too much operational work for one person, not enough revenue to hire three people.
So I built something different: a team of 7 AI agents that run my business operations 24/7 for $220/month.
After 14 weeks and 90 autonomous operating cycles, here are the real numbers — including the failures.
The Setup
Each agent specializes in one business function:
| Agent | Role | What It Does |
|---|---|---|
| Grove | CEO/Strategy | Sets priorities, coordinates agents, makes strategic calls |
| Burry | CFO/Finance | Tracks P&L from Zoho Books, flags expenses, questions ROI |
| Draper | CMO/Marketing | Content creation, campaign management, lead generation |
| Mariano | Sales | Pipeline management, outreach sequencing, follow-ups |
| Tars | CTO/Tech | Infrastructure monitoring, incident response, health checks |
| Drucker | Research | Competitive intel, market analysis, opportunity scanning |
| Warhol | Creative | Content production, brand voice, design direction |
Stack: Claude Code + MCP (Model Context Protocol) + Shared workspace + Task delegation system
Monthly cost: $220 ($100 Claude API + $20 server + $100 tooling)
The Numbers (14 Weeks)
| Metric | Value |
|---|---|
| Autonomous dispatch cycles | 90 |
| Emails sent | 432 |
| Unique contacts reached | 292 |
| Replies received | 23 (5.4% rate) |
| Total cost | $2,950 |
| Revenue | $0 |
Yes, $0 revenue. More on that below.
What Actually Works
1. Emergent Self-Correction
The most surprising finding: agents started catching each other's mistakes without being programmed to do so.
The finance agent questions the marketing agent's ROI claims. The research agent flags when data it previously provided has gone stale. The strategy agent reprioritizes when metrics shift unexpectedly.
This wasn't designed — it emerged from giving each agent clear domain ownership and visibility into the shared workspace.
2. Forced Forgetting > Persistent Memory
Counter-intuitive: agents with TTL-based context (auto-expire after N hours) made better coordination decisions than agents with access to full conversation history.
Why? Less noise. Fresher context. No anchoring to outdated information from weeks ago.
We use tiered TTL:
- Strategic decisions: 30-day TTL
- Business metrics: 7-day TTL
- Status updates: 24-hour TTL
3. Personality > Permissions
Telling an agent "you're a paranoid CFO who questions every expense" produced better financial oversight than restricting its API access.
Character constraints shape behavior more effectively than tool limitations in production.
4. $220/Month vs $10,000/Month
The equivalent human team for what these agents do:
- Marketing coordinator: ~$4,000/month
- Research assistant: ~$3,500/month
- Bookkeeper/admin: ~$2,500/month
- Total: ~$10,000/month
For routine operational work — research, data entry, email drafts, report generation, monitoring — the ROI math is clear.
What Doesn't Work
The $0 Revenue Problem
I spent 13 weeks marketing an AI operations system to... AI experts. Newsletter editors, tool builders, AI thought leaders.
They could build their own War Room in a weekend. I was selling hammers to carpenters.
The real market: Non-technical business operators with revenue who NEED AI operations but CAN'T build multi-agent systems themselves.
- Agency owners doing $500K-$5M drowning in ops
- E-commerce operators running $1M+ stores
- Professional services firms exploring AI
- Content businesses doing $100K+ revenue
These people see $2,500 as cheap compared to hiring an ops person ($50K+/year).
Trust Can't Be Cold-Emailed
432 outreach emails from an unknown AI sender = spam folder for most people. Cold email from an unfamiliar domain, no matter how personalized, cannot manufacture trust.
Community presence, published content, and social proof are prerequisites — not optional extras.
AI Can't Close Deals
Agents can research, draft, coordinate, and follow up. But the final handshake — the moment a prospect decides to pay — requires a human. Trust is analog.
The Architecture (For Builders)
Key design decisions:
- No central orchestrator — agents coordinate via shared workspace, not a master controller
- Human-in-the-loop for commitments — all external actions require approval
- TTL-based memory — context expires automatically, preventing stale data accumulation
- Personality-first agents — behavior shaped by character, not just permissions
What I'd Do Differently
- Target operators first, not builders. 13 weeks wasted on the wrong ICP.
- Community before outreach. Build trust in public before sending cold emails.
- Show the P&L, not the architecture. Business operators care about costs and outcomes, not MCP protocols.
- Start with one agent, prove value, add more. A 7-agent system is intimidating. One agent that saves 10 hours/week is compelling.
What's Next
The system works. The product is real. Now we need the right audience.
Pivoting to business operators: agency owners, e-commerce operators, and professional services firms who want AI-powered operations without the technical complexity.
War Room Setup-as-a-Service: Full 7-agent deployment on your infrastructure in 5 days. $2,500.
If you're drowning in operational tasks and curious whether AI agents could handle them — I'd love to hear what's eating your time.
https://warroom-landing.vercel.app
All data in this article is real. No demos. No simulations. 90 autonomous dispatch cycles over 14 weeks. The transparency is the product.
Top comments (0)