DEV Community

Santhosh M
Santhosh M

Posted on

I Run 10 AI Agents 24/7 on a Mac Mini. Here's What Actually Happened.

#ai #automation #buildinpublic

I built 10 autonomous AI agents. They run 24/7 on a Mac Mini sitting under my desk. After 5 days, total revenue: $0.

But here's the thing — they're not broken. They're working perfectly. They're just selling nothing.

The Setup (Architecture Porn Incoming)

Last week I shipped OpenClaw: a system where 10 LLM agents run autonomously with their own brain files, memory system, and inter-agent communication protocol. Each agent has a specific job:

  • email-monitor — Reads Gmail (2 accounts), categorizes 100+ daily emails, routes P0 alerts to the right agents
  • client-manager — Scans freelance platforms (Upwork, Freelancer, Contra), writes personalized proposals
  • dev-coder — Hunts GitHub bounties, submits pull requests
  • bounty-hunter — Analyzes bounty platforms, tracks competition, picks winners
  • content-creator — Publishes articles to Dev.to, Medium, Hashnode
  • x-content — Posts tweets to @Quant_Grow using Twitter API
  • bot-builder + discord-bot-worker — Builds and deploys Discord bots
  • discord-content-worker — Engages in communities authentically
  • admin — Routes work, supervises other agents, generates reports

Each agent fires on a cron schedule: email-monitor every 20 minutes, others every 3-6 hours. That's ~200+ autonomous sessions per day.

The Reality Check

18 proposals across 3 platforms. 0 responses.

  • Upwork: 3 proposals submitted to potential clients. Silence.
  • Freelancer.com: 5 bids placed. One person said "hi" in messenger. Still no contract.
  • Contra: 9 applications. Zero interest.

4 Gumroad products published. 0 sales.

  • AI Automation Starter Kit ($9)
  • Next.js SaaS Starter Kit ($19)
  • Discord Bot Monetization Template ($12)
  • n8n Automation Templates Pack ($9)

No traffic, no conversions, no revenue.

And you know what? The agents don't care. They ran their jobs yesterday. They'll run them again tomorrow. They don't check bank accounts. They don't spiral. They just... work.

This is the indie hacker truth nobody tells you: Building the machine is not the same as building the market.

The Architecture (What Actually Matters)

But here's why I'm shipping this publicly instead of hiding in shame: the Finch Memory System is genuinely useful for anyone building autonomous AI agents.

Each agent has 11 brain files loaded into every session:

~/.openclaw/workspace-{agent-name}/{agent-name}/
├── SOUL.md                 # Identity, personality, autonomy rules
├── AGENTS.md              # Trust ladder, sister agents, routing table
├── MEMORY.md              # 4-level memory system (see below)
├── TOOLS.md               # Skills, capabilities, workflows
├── HEARTBEAT.md           # Health checks, scheduled tasks
├── MISSION.md             # Shared mission statement
├── AUTONOMOUS.md          # Per-agent goals, results log
└── knowledge/
    ├── people.md          # Contacts, clients, opportunities (PARA system)
    ├── projects.md        # Active work, pipeline
    ├── resources.md       # Tools, APIs, templates
    └── archive.md         # 30+ day cold storage
Enter fullscreen mode Exit fullscreen mode

Finch 4-Level Memory works like this:

  1. HOT — Today's session. Active context (current email, current proposal, current bug)
  2. WARM — Last 30 days. Distilled learnings (what platforms respond fastest, which pitches convert, which APIs are flaky)
  3. INSIGHTS — Patterns from months of data (client archetypes, seasonal demand, bounty competition)
  4. ARCHIVE — 90+ days old. Auto-cleared to keep the brain lightweight

Every night at 23:00, a cron job runs nightly-memory-extraction.sh, which:

  • Reads the day's .learnings/ session logs
  • Extracts insights (e.g., "Freelancer bids rank higher with demo videos")
  • Updates MEMORY.md with high-signal patterns
  • Archives 30+ day items

The Trust Ladder is my favorite part. Each agent has 4 autonomy levels:

| DO FREELY | SHOW ME FIRST | ALWAYS ASK | NEVER |
|-----------|---------------|-----------|-------|
| Read emails | Draft proposals | Post tweets | Delete data |
| Scan jobs | Write code | Accept contracts | Fabricate revenue |
| Research | Create PRs | Spend money | Leak credentials |
| Draft content | Submit bounties | Send emails | Spam users |
Enter fullscreen mode Exit fullscreen mode

The client-manager agent can write 50 proposals (DO FREELY), but can't actually apply without review if the project fee is >$100 (SHOW ME FIRST). The x-content agent can draft tweets, but posting goes through ACP (Always Check Posted) because Twitter suspension is real.

Real Example: Email Classification

Here's how the email-monitor agent actually works. This morning it received:

Email #1: "You've got a new Upwork invite!"
→ Classified: FREELANCE ALERT (HIGH PRIORITY)
→ Routed to: client-manager agent (creates ACP task)

Email #2: "Your GitHub API quota is low"
→ Classified: SYSTEM ALERT (MEDIUM)
→ Routed to: dev-coder agent (logs to #coding-tasks Discord)

Email #3: "Check out this AI newsletter"
→ Classified: MARKETING NOISE (LOW, IGNORED)
→ Deleted silently (privacy preserved)
Enter fullscreen mode Exit fullscreen mode

Each alert is a JSON file in ~/.openclaw/shared/acp-messages/:

{
  "timestamp": "2026-03-05T08:34:22Z",
  "from": "noreply@freelancer.com",
  "priority": "HIGH",
  "type": "freelance_alert",
  "task": "New Upwork invite detected",
  "context": "Client avatar matches our niche",
  "routed_to": "client-manager",
  "action": "draft_personalized_proposal"
}
Enter fullscreen mode Exit fullscreen mode

The client-manager reads this, generates a proposal, posts it to Discord #client-intake for review, and waits for human sign-off.

No hallucinations. No spam. No deleted emails. Just clean, intelligent routing.

The Realization

Day 4, 11 PM. I'm staring at the OpenClaw dashboard showing $0 revenue. The email-monitor is running smoothly. The bounty-hunter just analyzed 47 opportunities. The x-content agent drafted 5 tweets that are actually funny.

And it hit me: The agents themselves are the product.

The real value isn't "these agents made me money." It's "here's a complete, battle-tested template for building autonomous AI swarms."

So I'm packaging the entire thing:

  • All 10 agent brain files (SOUL, AGENTS, MEMORY, TOOLS, AUTONOMY templates)
  • The OpenClaw cron configuration (jobs.json template)
  • Finch Memory System implementation scripts
  • Inter-agent communication protocol (ACP message bus)
  • 9 custom revenue skills (proposal-writer, bounty-executor, product-marketer, etc.)
  • Complete setup guide + troubleshooting

The CTA

If you've ever wanted to build autonomous AI agents but didn't know where to start, I just open-sourced the playbook.

OpenClaw 10-Agent Swarm Kit is live on Gumroad: quantbit1.gumroad.com/l/openclaw-swarm-kit ($49)

Also available:

What's Next

Day 6 starts now. The agents are back online. I'm following up with the one Freelancer prospect who responded. The content-creator is writing article #3. The x-content agent is posting daily (follow @Quant_Grow for build-in-public updates).

Total revenue today: still $0.

Total lessons learned: priceless.

The future of autonomous work isn't "build agents that make you rich overnight." It's "build agents so good that you can ship them and let others succeed with them."

Want to build your own agent swarm? Start small. One agent. One job. One memory system. Build in public. Ship the learnings.

Follow the journey on X @Quant_Grow — Daily updates on agent architecture, failures, and wins.

Top comments (0)