What it takes to run a fleet of AI agents: every piece of hardware and every service we use

How the fleet is organized

The fleet is set up like a small company. There are four kinds of jobs.

A chief of staff runs the founder's calendar, inboxes and queue. When something needs the founder, it goes through the chief of staff, so the founder sees one short list instead of 33 separate threads.

A chief engineer keeps every agent running. When a server, a scheduled job or a shared tool breaks, the chief engineer fixes it. Any problem that blocks more than one agent goes there too.

Project managers each own one business or project. A project manager decides what gets built for that business, writes the task, and is on the hook for it being done.

Specialists do one kind of work for everyone: security, compliance, design, video, search, content, email, social, paid ads, analytics, research, integrations, finance, and the skills the other agents use.

Here's how work moves between them. A project manager who needs something from a specialist files a task on that specialist's board, with a due date and a way to check it's done. The specialist accepts it or pushes back. Before certain steps, a project manager has to ask the right specialist first: security before anything that touches client data, compliance before any claim or price goes public, design before any logo or layout. Some decisions never belong to an agent at all. Spending money, anything you can't undo, and anything sent in the founder's name go to the founder.

1The computer, and how agents run on it

Everything starts with one machine, and it can't be your laptop. You close a laptop at night. The agents need to keep working after you stop.

The server (Hetzner). We rent one computer in a data center, and it never turns off. It's much bigger than a laptop: 36 processor cores and 125 GB of memory. Every agent lives on it. When we say "the fleet," this is where the fleet is.

Claude Code. This is the program each agent works inside. It can read files, run commands on the server and write code. You can think of it as the agent's desk and tools. The thinking comes from the AI model (Part 4), and Claude Code is where that thinking turns into work.

A window for each agent (tmux). Each agent gets its own window on the server that stays open all the time, 33 at once. tmux is the tool that keeps those windows alive when nobody is looking at them, so an agent keeps working after we log off.

Scheduled jobs (cron). Cron is the server's built-in alarm clock. You tell it "run this task at 6 AM every day" or "run this every 15 minutes," and it does, forever. We have about 370 of them. They send morning reports, check inboxes, restart the agents each night, and remind an agent when one of its tasks is past its due date. A lot of what looks like the agents "remembering" to do something is actually cron.

2Where the data and files live

Agents make and use a lot of information. It has to go somewhere safe, and it has to be easy to find again.

Databases (Supabase). A database is an organized store of records that an app can search, like a filing cabinet a program can read in a split second. Supabase hosts ours. The nice part is speed: when an agent builds a new app, it can set up a fresh database for it in a few minutes.

Shared memory. This is our own database, and it solves a problem every agent has. Each session starts with a blank head. Yesterday's decisions are gone unless they were saved somewhere. So when something matters, like a decision the founder made or a rule that changed, the agent writes a short row into it. The next morning, it reads those rows back before it does anything else.

File storage (Supabase and Vercel storage). The images, videos and voice files our apps use are stored here, close to the apps that show them.

Finished files (Dropbox). When a piece of work is done, the final file lands in one Dropbox folder tree. People can open it from anywhere without knowing anything about the server.

3Code

Most of what the agents build is code: websites, small apps, and the scripts that run the fleet itself. Code needs a home, and it needs a second pair of eyes.

Code hosting (GitHub). Git is a system that keeps every version of every file. If something breaks, you can see exactly what changed and roll it back. It also lets many agents work on the same code without wiping out each other's work. GitHub is where all of that is stored and shared.

Code review (Codex). Code review means someone else reads a change before it goes in. For us, that someone is OpenAI's Codex. Every change an agent makes is sent in for review, and it merges only after the review comes back clean.

Website and app hosting (Vercel). Every site and app we run lives on Vercel. Each change goes to a test copy first, and a person looks at it before it goes live. That test copy is where we catch problems while they're still cheap to fix.

Small services (Railway). One small media tool runs here. Some jobs fit better on a separate little host.

4The AI models

Each agent thinks with an AI model, and different models are better at different jobs. These are the models we use as of October 2026. They change often, and this list will change with them.

Claude (Anthropic). Claude Opus 5.5 runs every agent and does the judgment work: planning, deciding what to do next. Claude Sonnet takes work an agent hands off. Claude Haiku handles bulk reading, like going through long work logs. Claude Fable writes the prose people read.

OpenAI. GPT-6 Astra writes code. GPT-5.6 Terra is the fallback when it's busy. GPT-5.5 reviews code, through Codex.

Google Gemini. Gemini 2.5 Flash runs the chat and phone agents for one of our clients. Agents also call Gemini for some specific jobs and second opinions.

xAI Grok. Grok gives second opinions too, and it runs the live phone voice in Part 6.

OpenRouter. One account that reaches many other models. We use it for GLM-5.3 Flash, a cheap model for bulk work. When an agent wants to try a model we don't use every day, it goes through OpenRouter instead of us opening a new account.

Research search (Perplexity and Exa). These are search tools built for AI. An agent asks a question and gets answers with their sources, so it can check where a fact came from before it uses it.

5Integrations: how agents plug into other tools

An agent is only as useful as the tools it can reach. These are the three ways ours reach the outside world.

APIs. An API is a door a service leaves open for software. Instead of clicking around a website, an agent sends a request through the API and gets the answer back as data. Most of the tools in this article are reached this way, with no screen involved at all.

Composio. Composio connects our agents to Gmail, Google Calendar, Google Drive, LinkedIn and Meta. Each of those has its own API and its own rules, and Composio handles the connection so the agents can just use them.

MCP plug-ins (Model Context Protocol). MCP is a standard way to give an AI agent a new tool. Think of it like a plug that fits any socket: if a tool offers an MCP plug-in, any agent can pick it up and use it. We use MCP for the browser, for looking up software library docs, and for our own documents.

6Talking to people

Email, voice and phone are how our agents reach clients, partners and each other.

Email (AgentMail, Google Workspace, Resend and Instantly). We use four email tools, because email does four different jobs.

• AgentMail gives every agent its own inbox. An agent can send mail and read the replies, like anyone else on the team.
• Google Workspace runs the company mailboxes, calendars and shared documents that people use.
• Resend sends the emails our apps send on their own, like confirmations and receipts.
• Instantly sends outreach email. It spreads that mail across many inboxes, so no single inbox sends too much and gets flagged as spam.

Voice (ElevenLabs v4 and Grok Voice 3.0). ElevenLabs v4 turns a written script into a natural spoken voice, and we use it to narrate videos. Grok Voice 3.0 talks live, back and forth, as the conversation happens.

Texting and phone (Twilio). Twilio gives us text numbers and phone lines for clients.

7Making content

A lot of the fleet's work is content: posts, videos, images and articles like this one. These are the tools that make it, check it and get it out the door.

Images and video (fal.ai and Replicate). Each is one account that reaches many image and video models. When an agent needs a picture or a short clip, it picks the model that fits the job.

Stock footage and music (Pexels, Coverr and Epidemic Sound). Stock footage and licensed music for videos, so not everything has to be made from scratch.

Video editing (Descript). Descript edits video by editing the transcript. Cut a sentence from the text, and it's cut from the video.

Writing check (Pangram). Pangram reads a piece of writing and flags text that reads like it was written by AI. Everything we publish gets checked first. If a piece scores too high, it gets rewritten before anyone sees it.

Social scheduling (Postiz). Postiz queues posts for social media. Nothing goes into the queue until a person has approved it.

Social data (ScrapeCreators). ScrapeCreators reads public posts and profiles, so we can see what people are actually asking about and which posts are working.

8Running the work

Thirty-three agents need to know what to do next. They also need a quick yes from a person, and a way into websites that weren't built for them.

The board (Basecamp). This is where the agents work. Each business has its own project. Agents file tasks as cards, give them due dates, comment on them, and move them to Done when the work is finished and checked. If you want to know what the fleet did today, you read the board.

The approval page. We built our own page where the founder taps yes or no on posts and designs. It's faster than reading a long thread.

A logged-in browser. Some sites have no API. Business directory listings, some dashboards and account settings only work through a normal web page. For those, the agent uses an ordinary browser that's already signed in and clicks through the site the way a person would.

9Money, domains and measuring

The last part is the business side: getting paid, owning your web addresses and knowing whether any of it works.

Payments (Stripe). Stripe handles invoices, subscriptions and checkout.

Domains and DNS (Namecheap and DNS Made Easy). A domain is a website's name, like bluhook.com. DNS is the internet's address book: it points that name at the computer where the site lives. We buy domains through Namecheap and run the address book on DNS Made Easy.

Counting visitors (PostHog and Microsoft Clarity). PostHog shows which pages people visit and what they click. Clarity records how people move around a page, so we can see where they get stuck.

Search (Google Search Console and Bing Webmaster Tools). These are Google's and Bing's own reports on how our sites show up in search: which searches find us, and which pages have problems.

Where to start if you're building your own

You don't need all 37 on day one.

If you're starting today, begin with four: a server that never turns off, a database, GitHub with code review on every change, and one good model. That's enough for one agent to do useful work and for you to trust what it did. Add the rest the first time an agent gets stuck without it.

If you'd rather have a fleet like this built for you, tell us about it.