Bigger Box or Bigger Fleet?

I don't know exactly how we got here, but nowadays I can't seem to justify putting serious time into anything that isn't orchestration, architecture, design, and code review.

With a good enough harness and the right steering, it has become surprisingly easy to go from 0 to 80 on multiple projects in parallel. The 80 to 100 problem is a topic for a later write.

The constraint myself (and I'd imagine many others) are running into during this tokenmaxxing era is twofold:

1. Cost
2. Resources

This write-up is about my personal AI workflow, the limits I am hitting, and how I am thinking about scaling it up.

My Current Setup

For context, I currently have:

• Codex $20/month personal plan
• One relatively good desktop build
• A very old Lenovo IdeaPad 3 with an Intel i3
• A Fingerbot to physically turn my desktop on and off when I am away from home

Most of my personal work, compute, and inference happens on my desktop. The Lenovo is not really the machine doing the heavy lifting. I put a minimal Ubuntu install on it and use it more like a thin orchestrator. It is the machine I carry around, SSH from, and use as the entry point into my actual work environment.

When I am not home, the flow looks something like this:

Lenovo laptop
  ↓
Tailscale / SSH
  ↓
Home desktop
  ↓
Actual compute, agents, terminals, development work

If my desktop is off, I use the Fingerbot to press the power button remotely (there has to be a better way). Once the machine boots, my agents can SSH in and start working.

Yes, I realize this sounds like I am just rediscovering remote development.

But I kind of am...

The reason I am writing a whole thing about it is because I'm curious whether this is the "right way" to do things. In my head, how I want to continue to develop going into the future is simple: I need a single point of control for a workflow that increasingly wants to fan out into multiple agents, terminals, sandboxes, and tasks running in parallel.

Obviously, with rigorous slop clean-up involved.

The Lenovo Is the Orchestrator

Right now, my single point of truth is a Lenovo IdeaPad from 2020.

It has 8 GB of RAM.

It would be extremely ambitious to pretend this machine can comfortably spin up multiple Codex CLI sessions, multiple terminals, a browser, notes, docs, local services, and whatever else I want running at the same time. It definitely cannot run the kind of parallel workflow I actually want. It also DEFINITELY cannot run something like the Codex app, which is not even available on Linux as of this writing.

So in a sense, I kind of treat the desktop as my personal cloud.

The orchestrator does not need to be the strongest machine in the setup. It needs to be reliable and nice to use!

At least, that is the theory.

It must also be mentioned that although I am rigorously trying to tie AI into my workflows, I am still fully responsible for all of its output.

I feel like code ownership is the most important part of this era, as described in the shovel write, so when reading the next section, keep in mind that I am assuming that I will be able to maintain full code ownership under the potential increased rate limits.

The First Wall: Cost

I am currently on the $20 Codex plan. In theory, you might ask:

How are you getting so much usage out of a $20 plan while running multiple agents in parallel?

Well friend - I am not, really.

It is difficult to run Codex the way I want to for more than a short stretch before I start feeling the limits. When your workflow depends on parallelism, the constraint shows up FAST, especially when Codex gets ambitious and starts dividing up work like crazy.

That is where the $20 to $100 jump starts to look less absurd to me.

At first, $100/month feels like a lot. But if the tool is meaningfully increasing the amount of work I can get done, then it starts looking more like infrastructure cost than a subscription.

And as we will talk about, $100 a month to 5x my work (generous estimate, but you get the idea), compared to a MacBook with the same price tag as a used 2017 Volkswagen Tiguan, seems well worth it.

Now the question seems obvious:

Is my current workflow bottlenecked by usage limits, or resources?

Well, yes, but in this hypothetical world where I have 5x the rate limits on my plan, we run into...

The Second Wall: Resources

If I increase my usage, can my current hardware setup actually support the way I want to work?

"Bigger box vs bigger fleet" should start to make a little more sense as the title of the write.

Do I invest in one stronger machine that can do everything? Or do I build a small fleet of cheaper machines that can each take on a piece of the workload?

There is a world where you do both. You have one very powerful primary machine and a fleet of smaller boxes for sandboxes, services, experiments, and agents.

But for a normal person trying to balance AI subscription costs, hardware costs, and actual day-to-day usefulness, it's kind of one or the other, ESPECIALLY with the hike in computer hardware pricing in recent months.

Bigger Box

The bigger box approach is simple.

Buy one powerful machine. Make it the center of everything. Give it enough CPU, RAM, storage, and GPU headroom to handle whatever you throw at it.

The appeal is obvious:

• Less complexity
• Fewer moving parts
• Easier maintenance
• One environment to configure
• One machine to secure
• Better for large local models
• Better for resource-heavy development

This is the dream version of the setup to some: one box for all!

In that world, you just buy the best MacBook Pro, desktop workstation, or custom build you can afford and call it a day. Everything runs there. Your agents, editors, terminals, services, local models, containers, and sandboxes all live on the same powerful machine.

It is clean...

It is also expensive...

You can spend an absurd amount of money chasing the perfect machine. The fully maxed-out laptop fantasy is always there, just at a price tag of $10,000! Just buy the biggest, fastest thing and stop thinking about infrastructure!

But that feels like a trap too. There will always be a newer chip, a better GPU, more RAM, faster storage, and heavier workloads. Today it is the M5 Max. Tomorrow it is the M6, then the M7, and so on.

Vertical scaling is seductive because it simplifies the system. But it also concentrates your entire workflow into one expensive box.

However, as a brilliant engineer once told me:

A single point of failure is a single point of tracing and recovery...

Bigger Fleet

Instead of buying one monster machine, you buy or reuse multiple cheaper machines and connect them together.

You might have:

• A laptop as the orchestrator
• A desktop as the main worker
• A mini PC for sandboxes
• Another cheap box for background services
• A NAS or storage box
• A machine dedicated to local model experiments
• Disposable environments for agents to work inside

You can isolate workloads, give agents their own sandboxes, and avoid letting one task eat all the resources on your main machine. You can experiment without risking your primary environment and build a little personal compute cluster, held together by SSH, TMux, Tailscale, and blipcoard :)

BUT, more machines means more updates, more networking, more secrets, more failure points, more weird bugs, and more things to forget you configured six months ago.

But maybe that is the direction personal AI workflows are going anyway, since your agent could just handle all of that, theoretically.

Open-Weight Models Complicate the Decision

There is another piece here too: open-weight models.

Right now, I do not think open-weight models are good enough for every heavy workload I care about, especially compared to the best hosted models.

But it would be shortsighted to assume that will always be true, especially as I grow into a better engineer. In a perfect world, I won't need more than something like DeepSeek V3 Flash to help me execute faster.

If local models become increasingly useful, then the bigger box argument gets stronger.

Horizontal scaling is useful for parallel sandboxes and distributed work, but it does not help as much if what you really need is one machine powerful enough to run a large model well. For that, you need serious local resources.

GPU, memory, bandwidth, thermals: all the annoying physical stuff comes back into the picture.

So, what's really future-proof?

Do you invest in a fleet because agentic workflows want parallelism?

Or do you invest in one bigger box because local inference wants concentrated power?

The answer is probably both or neither to be honest, but for the former, we are not all Peter Steinberger and for the latter, NOTHING is future-proof when things are moving so fast in our world right now.

My Choice

I am obviously talking slightly out of my league here. I am not building a real cluster. I am not spending $10,000 on hardware. I am just trying to decide whether moving to the $100 Codex plan and upgrading part of my setup would meaningfully improve my workflow.

Even then, I would not want to spend more than around $2,000 upgrading either the fleet or the box.

Right now, I am leaning toward replacing the Lenovo orchestrator.

Part of that is practical. It is old, underpowered, and not especially pleasant to use.

Part of it is also just that I want to chase a new, shiny screen...

But I am trying to be honest with myself about what the laptop actually needs to do. If the desktop remains the real worker, then the laptop does not need to be a monster, even though knowing I could have 64-128 GB of RAM is awesome.

Not everyone's workflow needs to head in this direction. Most people probably should not care.

But if you are trying to use AI tools as an actual extension of your development process, not just as a chatbot in a browser, then these decisions will probably come up fast if you are consistently progressing each and every day.

In conclusion, keep tokenmaxxing responsibly during the golden age of subsidization :D