Three Terminals, Zero Myth

Three Terminals at Once

For several hours today I had three Claude Code terminals running simultaneously. And that's not counting all the autonomous agents I have performing various tasks in the background. I've been saying it's compounding every day, and today was the day I actually felt it.

I know there are people out there handling way more terminals than me. People buying bigger monitors just to fit them all on screen. I'm not trying to brag. But here's what I want to say to anyone who hears those stories and thinks "sure you did, but do you do that daily? All day?" The answer is yes. These people are 100% accurate. Don't automatically doubt it. They're either doing it or they know somebody who is and they want to sound cool. It's not a myth. I got my first real taste of it today.

Is It the Engineer or the Tooling?

The productivity drop from context switching is real. And in a way, the fact that I can do this kind of tells on me. It says something about the complexity of the tasks I'm working on. But it also says something about how far AI and the ecosystem around AI has advanced. Both matter. It's not just about how smart the models are. It's how effective the tooling around the model is.

We take for granted how much tooling is actually guiding these models to be effective, productive, and reliable. If you go back to the earliest implementations of ChatGPT as a baseline and compare that to all the agentic tooling we wrap our models in today, the difference is massive. If you were to go back to a raw form of interacting directly with the model, you'd still get a lot of weird, unusable stuff. Maybe at its core it hallucinates less now, but it's still going to be problematic getting useful results out of it without the layers of prompting on top, the reasoning loops. That's not all baked into the model. A lot of it is the tooling.

A New Kind of Stack

You should always be prepared to add your own layering. There's a new kind of stack forming. I want to be cautious about claiming to understand the lower levels of it, but as I build my own tooling, work with my own team of agents, delegate responsibilities, and configure different specialties, I actually feel like I have an intuition about what's happening on the lower levels of the stack, right above the raw model.

Where my understanding starts to come apart is at the weights. The actual training, what goes into shaping the model at that level. I'd love to dive into that at some point. But right now there's so much to learn with the tooling and orchestration of agents, writing rules, writing skills, building tools. There's so much in that part of the ecosystem that I don't see a path where I'm working with training anytime soon.

But as far as all the tooling? That's the space I'm hands-on with right now, and I'm getting amazing results. I mentioned recently playing around with building an agent as a tool. That has been life-changing.

Strategic Client Overlap

A factor in how I'm able to pull this off is a strategic freelance decision. I'm a freelancer, and to some degree I factor in overlap when I'm taking on new clients. Overlap in frameworks, stacks, the kinds of problems they face. If I'm working with multiple tradespeople, they're probably all going to use Housecall Pro or something like it. If I'm working on custom SaaS products, there's going to be overlap.

I'm not going to recommend a tool that isn't the best tool for the job. But if I can choose between potential clients where the best tool for the job is Next.js deployed to Vercel, I'm going to do that. Because while I'm experiencing certain things for one client, I can apply them to another. These aren't theoretical scenarios. This is what I'm really dealing with right now.

I've got multiple clients I'm doing digital marketing work for, and it made sense for all of them to use Next.js deployed to Vercel. So now I'm running Lighthouse scans trying to get my LCP (Largest Contentful Paint) numbers in line so Google considers these sites good. And my LCP numbers were atrocious. If I discover that an older React version makes it impossible to hit those scores, I'm going to know to go to my other clients and say, hey, we need to upgrade your React version. The new standard has been set.

Maybe the newer React client-side JavaScript bundle is smaller. Maybe it's faster. Maybe there are features that behave differently and we need to refactor how we're handling the render strategy. You need that LCP under about 300 milliseconds. And if you've got a large hero image above the fold, which is a very popular design that users love, your LCP is easily going to be over a second or two. As soon as that happens your Lighthouse score drops below 90, and Google considers your site less than good. It doesn't matter if you have the most beautiful hero banner the world has ever seen if nobody sees it. If you're invisible, it doesn't matter how pretty it is.

There's an economy of scale in picking clients with related tech stacks. What you learn solving problems for one directly benefits the others.

What Does a Dry Run of Autonomy Look Like?

To address the concern that maybe I'm not doing novel enough things if I can context switch between three terminals, let's talk about what's automated in the background. The blog publishing pipeline. The SEO audit pipeline. The Upwork scraper. The daily stand-up meetings where odd tasks that don't have a pipeline are happening autonomously. All of these were once tasks I was doing manually.

Part of what I'm doing while I bounce between these three terminals is a dry run of autonomy. Anything I can toggle between like this becomes a test case. If I, as a human, can manage the context switching between three workstreams, then there's a higher probability I'm inching closer to full autonomy for each one. It gives me the ability to ask: am I being forced to lock in and give all my focus to one of you? If so, that one isn't close to automation yet. Pay closer attention. But if context switching is easy for the entire day, that means it's time to put some guardrails on and see how it does on its own.