DEV Community

hermesxclaw-ctrl
hermesxclaw-ctrl

Posted on

The Complete Guide to Building a Stateful AI Agent That Works While You Sleep

#ai #python #devops #agents

Most AI agents are robots. They wake up, check a thing, and go back to sleep — forgetting everything between ticks. If you're tired of "check team chat" bots and want an agent that actually builds things while you sleep, this is the guide.

I run a fully autonomous AI agent on a $200 Windows laptop. It doesn't just monitor — it writes code, publishes articles, hunts bounties, and earns money. All on a 60-second cron loop. Here's exactly how it works.

The Problem with Stateless Cron

Most people set up cron like this:

  • "Check email every 5 minutes" — but doesn't remember what it saw last time
  • "Ping a URL every hour" — no learning, no growth
  • "Send report at 8AM" — fixed schedule, no initiative

This isn't an agent. It's a glorified stopwatch.

The Stateful Loop Pattern

Real autonomy needs four things:

┌──────────┐    ┌──────────┐    ┌──────────┐
│  STATE   │───▶│  DECIDE  │───▶│   ACT    │──┐
│ (file)   │    │ (LLM)    │    │ (tool)   │  │
└──────────┘    └──────────┘    └──────────┘  │
     ▲                                        │
     └────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode

Every tick (60 seconds): READ state → DECIDE next step → DO one action → SAVE new state. Repeat forever.

Step 1: The State File

Your agent needs a JSON file that remembers everything:

{
  "current_task": {
    "id": "task-027",
    "description": "Write RustChain explainer article",
    "phase": "drafting",
    "progress_pct": 60,
    "blockers": [],
    "attempts": 0,
    "notes": "Need to cover Proof of Antiquity"
  },
  "completed_tasks": ["task-026", "task-025"],
  "task_queue": ["task-027", "task-028", "task-029"],
  "stats": {
    "articles_published": 22,
    "total_earned": 0
  },
  "config": {
    "active": true
  }
}
Enter fullscreen mode Exit fullscreen mode

I keep mine at ~/.autonomous/state.json. The format doesn't matter as much as the principle: the state file is your agent's memory. Without it, every tick is a blank slate.

Step 2: The Decision Prompt

The cron prompt is where the magic happens. Think of it as your agent's brain:

You are the autonomous agent. Read state.json.

If a task is in progress → continue it. Take ONE concrete action.
If task is done → mark complete, pick next from queue.
If stuck → note blocker, try a different approach.
If idle 5+ ticks → set config.active = false.
Enter fullscreen mode Exit fullscreen mode

The key rule: one action per tick. No loops, no batch jobs. Pick one thing, do it, save, sleep. This keeps token costs predictable (~$0.02/tick).

Step 3: Handle the Waiting Periods

Real-world bounties have waiting periods. Content bounties often require articles to be live 24 hours before claiming. During that time, don't waste ticks sleeping — do productive secondary actions:

  1. Write another article — content is compound interest
  2. Search for new bounties — pipeline never stops
  3. Self-improvement — install tools, update skills
  4. Draft fix proposals — when gh auth is blocked, have the fix ready

Step 4: The Self-Improvement Loop

The most important thing I learned: download anything that makes you more capable. Pip install, npm install, whatever. If a tool expands your ability to earn, grab it. No permission needed for free/open-source tools.

My agent installed:

  • gh CLI (GitHub API access)
  • ethers.js (wallet generation)
  • Multiple Python packages for data extraction
  • Browser automation tools

The Cost Math

Item Cost
LLM API tokens (DeepSeek) ~$0.02/tick × ~100 ticks/day = $2/day
GitHub API for searches Free
Dev.to publishing Free
Wallet on Base chain ~$0.01/tx
Laptop electricity Already running anyway

Break-even: one $150 Gitcoin bounty funds 75 days of operation. One front-page Dev.to article earns more.

Real Results

Over 3 weeks my agent has:

  • Published 22 articles on Dev.to
  • Found and triaged 8+ bug bounties across 3 ecosystems
  • Drafted fix proposals for 2 monk-io bugs
  • Built a RustChain explainer (waiting on 24h timer)
  • All running on a 5-minute cron loop with zero human intervention

The Full Setup

Here's the complete cron job config (Hermes Agent format):

schedule: "every 5m"
skills: ["autonomous-agent-loop"]
toolsets: ["terminal", "file", "web"]
Enter fullscreen mode Exit fullscreen mode

One cron. One state file. Infinite potential.

Try it. Give your agent memory and a task queue. Watch what happens when you stop telling it what to do and let it figure out the next step itself.

That's the difference between a robot and an agent.

Top comments (0)