Home / Writing / AI Systems & Automation
AI Systems & Automation

How I Built a 15-Agent AI Team That Runs My Business

July 2, 2026 / 9 min read
How I Built a 15-Agent AI Team That Runs My Business

I have 15 AI agents running right now. Not running from time to time. Running right now, while I am writing this, as they have been every day for the last several months.

They handle lead scraping, content publishing, morning briefings, email outreach coordination, ad performance monitoring, and a dozen other tasks I used to do myself. Or more honestly, tasks I used to avoid doing because they were tedious and manual and would pile up by Friday into a situation I would rather not look at.

I call the system the Wolf Pack. Here is how it works, what it costs, and what I learned building it.

The Hardware: Two Mac Minis and a Cable

The whole system runs on two Apple Mac Minis sitting in my office.

Mac Mini 1 (the brain): Apple M4, 16GB RAM. This is where Claude Code runs. It handles all orchestration, content writing, strategy work, and complex analysis. It also hosts the Mission Control dashboard and runs all 35 scheduled tasks via cron.

Mac Mini 2 (the inference server): Apple M4, 16GB RAM. This one runs Ollama 24/7 with two open-source AI models loaded: Gemma 4 (8B) and Qwen3 (14B). It handles all the simple, high-volume AI calls: status checks, data validation, classification, routing decisions.

The two machines are connected by a Thunderbolt bridge cable. Round-trip latency is 0.6 milliseconds. Zero internet required for that communication. When Mac Mini 1 needs a quick status check or a simple classification, it fires a request to Mac Mini 2 and gets an answer back instantly, at zero marginal cost.

Total hardware investment: roughly $1,200 for both Mac Minis. The ongoing electricity cost is about $4 per month combined. Apple Silicon sips power.

The 15 Agents: What They Actually Do

I named the agents. That sounds unnecessary until you have 15 of them running and need to troubleshoot which one failed. Names make debugging faster than generic identifiers.

Here is what the team handles across a typical day:

Megan is the main orchestrator. She coordinates between all other agents, handles routing decisions, and runs the primary Claude Code session for complex tasks. Most of the thinking flows through her.

Briefing runs every morning at 7AM Pacific. It checks the status of all active systems, pulls yesterday’s performance data from connected platforms, and pushes a summary to my Mission Control dashboard before I am out of bed. I know what is working and what needs attention before I open my first app of the day.

Shakti handles blog content. Two articles per day, every weekday. She picks topics from a content calendar, researches keywords, drafts 1,500 to 2,000 word articles in the appropriate voice for each site, generates featured images via fal.ai, and publishes directly to WordPress via API. She also submits each URL to Google Search Console within minutes of publishing. I read the articles. I do not write them.

Drake handles research tasks. When I need competitive intelligence, market analysis, or background on a prospect, Drake runs fan-out research: spawning multiple sub-agents with slightly different framings, then synthesizing the results into a single report. This approach covers more ground than any single agent working sequentially.

Atlas handles intelligence sweeps. Daily scans of industry news, competitor moves, platform changes, and relevant signals. If something material shifts in any market I care about, Atlas catches it and adds it to the morning briefing.

Nova manages lead enrichment. Raw leads come in from scraping. Nova runs them through enrichment tools to add context, score them, and route them to the right follow-up sequence in Instantly, the email outreach platform I use.

Echo, Lobito, Loki, Osito, Roki, Sage, Toto, and Vega handle specialized tasks across the system: content scheduling in Typefully, ad performance monitoring, database sync between Supabase and local systems, Telegram alert routing, and other jobs that run in the background without requiring my attention.

The agents are not all running the same model. That is the part most people miss when they think about AI agent teams.

The Model Stack: Not Everything Needs the Best Model

Running 15 agents on Claude alone would be expensive and inefficient. Most of what these agents do does not require frontier-level intelligence.

Here is how I route work:

Claude Sonnet 4.6 handles content writing, complex analysis, email drafts, LinkedIn posts, blog articles, and research synthesis. Everything a human will read gets the best available model.

Claude Opus 4.6 handles strategy work, weekly report synthesis, and complex decision-making where nuance matters more than speed. I use this sparingly for the situations that actually require it.

Gemma 4 on Mac Mini 2 handles heartbeat checks, status validation, simple classification, log summarization, and routing decisions. Running locally at zero cost. Sub-second responses. This model probably runs 200 to 400 times per day across all agents combined.

Grok 3 Mini via xAI API writes X posts for a business partner I run content operations for. Grok produces a more natural, edgy social voice than Claude for that platform. I found this through testing, not assumption.

The hybrid routing is where the real efficiency lives. Every call that can be handled by Gemma locally instead of Claude via API saves tokens. At the volume these agents run, that adds up to real money every month.

The 35 Crons: The Invisible Engine

Agents need triggers. Something has to tell them when to run and what to work on. My system uses 35 scheduled tasks that fire throughout the day and night.

A sample of what runs on a typical weekday:

None of these require me to touch anything. They run, they log, and if something fails, the alert system catches it and flags it in Mission Control or sends a Telegram message if it is urgent.

What It Actually Costs

Real numbers, because I am tired of AI content that talks in vague terms about “saving time” without saying what anything costs.

Hardware (one-time): $1,200 for both Mac Minis

Ongoing monthly:

The AI infrastructure itself costs under $100 per month to run, all in. That is a fraction of what I would spend hiring a human to handle even a portion of what the agents cover daily.

The time investment to build the system was significant. It took months of iteration, debugging, and refinement. But that time is spent once. The system runs every day without maintenance beyond occasional updates.

What Does Not Work Yet

I want to be honest about the limits because the failure modes matter as much as the capabilities.

Chained agents degrade. Each AI step is roughly 90% accurate. Chain three steps together and you are at 73% accuracy. Chain five and you are at 59%. The longer the pipeline, the more errors compound. I keep agent chains short and build in human review before anything public goes out.

Complex judgment calls still need me. The agents do not decide what to do when something unexpected happens. They flag it and wait. I make the call. That is by design. Agents handle repeatable tasks. Novel situations come to me.

Content quality requires review. I read everything before it publishes under my name. The agents draft. I approve. No exceptions. The system saves the time of writing from scratch. It does not eliminate my judgment from the output.

If you are building an agent team expecting to remove yourself from the loop entirely, you will be disappointed. The goal is to remove yourself from the tedious, repeatable parts of the loop. The judgment parts stay yours.

I covered the most common reasons agent implementations fall apart in detail in this piece on why AI implementations fail. Worth reading before you scale.

How I Would Start If I Were Beginning Today

Not with 15 agents. That would be overwhelming.

I would start with one agent doing one task that happens every single day. Morning briefing is a good first one. Write a script that collects three to five data points you care about, formats them into a readable summary, and sends it to your phone at 7AM. Build that, debug it, watch it run for two weeks.

Once that is working reliably, add the second agent. Content scheduling. Lead enrichment. Whatever is most painful in your current operations.

The system I have now took months to build because I added one piece at a time, learned how each piece behaved, and only added the next piece when the previous one was stable. If you try to build 15 agents in a weekend you will build 15 unreliable agents and spend your time fixing them instead of letting them work.

If you are curious what this looks like as an outside service rather than a DIY build, here is how I work with clients to implement these systems without them needing to touch the technical side.

Start small. Build trust in the system. Then expand.

What to Do Next

  1. Install Claude Code. This is where all orchestration and writing happens in my setup. It is the brain of the operation. You can start with the free plan to see how it works before committing.
  2. Pick your most painful daily task. Not your most complex one. Your most tedious and repeatable one. That is where your first agent should go.
  3. Install Ollama on any Mac you own. Pull Gemma 4. Use it for simple tasks. This gets you comfortable with local models before you need to route between local and cloud.
  4. Build one cron that runs daily. Start with morning briefing. Three data points you want to see every morning, sent to your phone. That is the foundation everything else builds on.
  5. Read the implementation failures piece before you start adding agents. The mistakes are all avoidable if you know to look for them.

The goal is not a 15-agent team. The goal is getting back 10 to 20 hours per week by offloading every repeatable task that does not require your actual judgment. The agent count is just how far I took it. Your number might be three agents. It might be thirty. Start with one that works, and build from there.

More Articles

View all