Home / Writing / Uncategorized
Uncategorized

How I Built a 15-Agent AI Team That Runs While I Sleep

July 30, 2026 / 11 min read
How I Built a 15-Agent AI Team That Runs While I Sleep

I’m not a programmer. I want to be clear about that upfront.

I can’t write Python from scratch. I don’t know what a Docker container is in any practical sense. And until about 18 months ago, my most sophisticated “automation” was a shared Google Sheet with a few formulas in it.

Today I have 15 AI agents running on two Mac Minis in my home office. They write blog posts, analyze ad performance, scrape leads, send me daily briefings, monitor my business health, and schedule content across three platforms. They run 24 hours a day, seven days a week, while I sleep.

This article is about how that happened, what it actually costs, what breaks regularly, and whether it makes sense for your business.

What This Is Not

It’s not a plug-and-play solution. There’s no software you buy and install that does what I’m about to describe. Every piece of this was built custom, one agent at a time, over about eight months.

It’s not “10x your business in 90 days” content. I’m going to give you real numbers, real failure rates, and real costs. Some of what I built doesn’t work as well as I expected. I’ll tell you that too.

And it’s not right for every business. The honest answer is that most small businesses should not try to build what I built. I’ll help you figure out if you’re the exception at the end.

The Hardware Setup

Everything runs on two Mac Minis. Total hardware cost: around $2,400.

Mac Mini 1 (the brain): Apple M4, 16GB RAM. This is the orchestration layer. It runs all scheduled tasks, hosts the monitoring dashboard, manages the content pipeline, and decides what to delegate. It’s connected via Cloudflare Tunnel so I can reach it securely from anywhere.

Mac Mini 2 (the inference server): Also M4, 16GB RAM. This one runs Ollama 24/7 with two local AI models: Gemma 4 (8 billion parameters) and Qwen3 (14 billion parameters). These are free, open-source models that run entirely on the Mac Mini hardware — zero API cost per query. The two machines connect via a Thunderbolt cable with 0.6 millisecond latency. Faster than most internet connections.

Why two machines? Because I needed the brain to always stay responsive for web requests and task management, while the inference server handled heavy AI processing without competing for the same CPU and memory. I made this call after six weeks of trying to run everything on one box and watching it slow to a crawl during heavy processing windows.

The Software Stack

Here’s what runs on top of the hardware:

The 15 Agents and What They Actually Do

I use the word “agent” loosely. Some of these are sophisticated multi-step reasoning chains where Claude plans, executes, and evaluates results across a dozen steps. Some are simple Python scripts on a timer. All of them run autonomously, without me touching them.

Content Agents:

  1. LinkedIn Content Writer — Writes 5 LinkedIn posts per day for my client Devon. Researches current events, matches his voice, formats for LinkedIn’s engagement patterns. Posts automatically via a scheduling queue.
  2. Blog Publisher — Writes and publishes 2 blog articles per day to three different WordPress sites. Generates featured images via fal.ai, uploads them, submits URLs to Google Search Console. Running right now, including this article.
  3. Devon X Writer — Writes 5 tweets per day for Devon’s X account, queues them via Typefully API. Uses Grok 3 Mini (xAI’s model) for the writing because it matches Devon’s voice better than Claude does. Claude is too polished for what works on X.
  4. TrueWater Blog Publisher — Dedicated agent for my water treatment business in Idaho. Researches local Boise-area news, finds relevant angles, writes and publishes local SEO articles targeting home owners in specific Treasure Valley cities.

Intelligence Agents:

  1. Ad Analyzer — Pulls creative performance data from Meta every morning. Identifies what’s working, what’s dying, surfaces patterns across campaigns. Logs findings to the Mission Control dashboard.
  2. Email Monitor — Watches a specific inbox for leads and important messages. Classifies them, routes them to the right queue, logs entries to Supabase.
  3. Devon News Scout — Searches the internet daily for news relevant to Devon’s business and audience. Formats discoveries for content ideation. Gets fed into the content writer’s context.
  4. LinkedIn Watchdog — Monitors Devon’s LinkedIn ad campaigns for performance drops, budget pacing issues, or audience delivery problems. Alerts me when something needs attention.

Operations Agents:

  1. Daily Briefing — Every morning at 7am, synthesizes overnight activity across all systems and posts a structured summary to Mission Control. Includes the day’s 3-5 tasks that actually require me — the things only I can do.
  2. Lead Enrichment — Takes raw leads from scraping jobs and enriches them with verified contact data via Prospeo’s API. Writes enriched records back to Supabase.
  3. Data Sync — Keeps Google Sheets, Supabase, and HubSpot in sync. Runs every 30 minutes. I don’t think about it anymore.
  4. Heartbeat Monitor — Checks system health every hour. Uses the local Gemma model — not Claude, zero cost — to classify each check and alert me only when something is actually broken.
  5. Claude Watchdog — Monitors all running Claude processes. Kills any that have been running longer than 2 hours. Stuck Claude runs are a real problem: the model can get into a reasoning loop and run for hours on a task that should take 10 minutes. Without this agent, I’d hit API limits on stuck processes regularly.
  6. Cloudflare Tunnel Monitor — Makes sure the tunnel connecting my Mac Mini to the outside world stays connected and healthy.
  7. Telegram Bridge — Receives messages from me on Telegram and routes them to the right agent, logs them as tasks, or triggers immediate actions. My interface to the whole system while I’m mobile.

What Runs While I Sleep

Here’s what a typical overnight sequence looks like:

11pm: LinkedIn content writer fires. Writes tomorrow’s 5 posts for Devon. Saves them as drafts in the scheduling queue.

11:30pm: Devon X writer fires. Writes 5 tweets, queues them in Typefully with the right posting times spread across the next day.

2am: Blog publisher fires. Writes and publishes two articles. Generates images, uploads to WordPress, submits URLs to Google Search Console.

4am: Ad analyzer pulls overnight Meta performance data. Logs results to the database. Flags anything unusual.

6am: TrueWater blog publisher fires. Researches Idaho water news, writes a local SEO article, publishes it with a generated image.

7am: Daily briefing agent synthesizes everything from overnight. Posts a structured summary to Mission Control. Sends me a Telegram message with the day’s quest — 3 to 5 tasks that require a human.

By the time I wake up, the operational work for the day is already done. My job is to do the things only I can do: show up to calls, review and approve content before it posts to social, record video, make phone calls.

What It Actually Costs

I see people online claiming AI automation costs almost nothing. That’s true for hobby projects. It’s not true for a production system running 24/7 at business scale.

Here’s my real monthly breakdown:

Total monthly operating cost: approximately $330-340/month.

Hardware was $2,400 one-time. At $330/month operating cost, I’ll hit break-even on hardware in about 7 months from what I’m saving in outsourced work — assuming the system actually works, which it mostly does.

What It Gets Wrong

Any honest account of an AI system needs to include the failures. Here are the real ones:

Compound errors in long chains. Every AI step is accurate roughly 85-90% of the time. Chain 5 steps together and you’re at 59% overall accuracy. The longer the pipeline, the more likely something weird gets into the output. I review social content before it goes out. Blog content I review in batches, not in real time — which means occasionally something publishes that’s slightly off. It’s not catastrophic, but it’s not zero.

Stuck processes. Claude occasionally gets into a reasoning loop and runs for hours on a task that should take 10 minutes. The Claude Watchdog kills these before they burn through limits, but diagnosing why they happened takes time I don’t always have.

API changes breaking things silently. Three times in the past year, a platform changed something in their API that broke my agents. Meta, LinkedIn, a WordPress hosting provider. Each time it took 2-6 hours to diagnose and fix — and the agents just silently fail until you notice. You need a monitoring layer, or you’ll discover failures days later when you realize no content was posted.

Voice drift. AI writing quality degrades over long sessions with lots of context. The eighth article in a batch sounds less like the intended voice than the second. I’ve mitigated this by keeping each agent run short and giving each one a fresh context. But it’s still something to watch.

Setup time. Building this took about eight months of evening and weekend work, with AI doing all the coding. If I’d paid a developer to build it, I’d estimate $40,000-60,000. The AI-assisted version cost me my time plus the API costs. Your tolerance for ambiguity, debugging, and iteration will determine how long it takes you.

Is This Right for Your Business?

Probably not — and I mean that seriously. Most small businesses should not try to build what I built.

This setup makes sense if:

It does not make sense if:

That last one is the most common mistake I see: businesses trying to automate processes they haven’t figured out manually. The AI will faithfully execute the wrong process, faster. You need to know what good looks like before you hand the keys to an agent.

Start smaller. One automation. One agent. Run it for 30 days. Fix what breaks. Then decide if you want to build the next one.

What to Do Next

  1. Pick your highest-frequency repetitive task. What do you or your team do more than 5 times per week that follows a consistent pattern? That’s your first automation candidate.
  2. Document the process before touching AI. Write down every step, every decision point, every edge case. If you can’t write it down clearly enough for a new employee to follow it, an AI agent can’t execute it reliably either.
  3. Start with Claude Code or a similar AI assistant. You don’t need to know how to code. You need to know what you want. Describe the task in plain English and let the AI figure out the implementation. I’ve built everything described in this article without writing a line of code myself.
  4. Build one thing. Run it for 30 days. Resist the urge to build everything at once. One working agent that runs reliably is worth more than five agents that break and require constant babysitting.
  5. Talk to someone who’s done it. I consult with business owners on AI implementation. If you want a direct conversation about whether this makes sense for your specific situation, reach out here. I don’t do discovery calls to pitch you. I do them to give you an honest answer about whether AI automation is actually the right move for your business right now.

More Articles

View all