I haven't opened my code editor in three months. I don't miss it.

The moment I noticed

I'm not a VS Code user. I'm a Sublime Text user, and I have been for a long time. This week it struck me that I haven't opened it in about three months.

Nobody told me to stop. I didn't decide to stop. It just stopped being the first thing I open in the morning, and then it stopped being something I open at all. When I looked for the moment it changed, I found it: it's when I moved all my coding agents off my laptop and onto a server.

I didn't quit my editor. My work moved somewhere the editor isn't.

Why my agents moved to a server

Three reasons, and none of them was about the editor.

My laptop ran out of memory. I work on a MacBook Air M2 with 24 GB of RAM. That isn't enough to run all my agents at once. So I took a Hostinger VPS with 32 GB of RAM and 8 virtual CPUs, their biggest plan, and gave the agents a machine of their own.

The agents were fighting my own work. While they worked, I worked too: content, and video editing. Video editing and a fleet of agents on the same laptop compete for everything, and both lose.

I'm always on the go. Closing the lid shouldn't stop the work. I wanted agents that run 24/7, whether my laptop is open, shut or in a bag.

Why not the hosted versions, Claude Code or Codex running in the vendor's cloud? Because my agents work in secured environments that need their own logins and configuration. That's easy on a box I control, and neither easy nor scalable on someone else's cloud.

How I work now

Getting started on a VPS is simple. I connect over SSH, and tmux keeps each session alive between connections. Then I started working from my Android phone, with Moshi, a phone terminal that speaks Mosh, a protocol that holds up far better than SSH when a phone drops between Wi-Fi and mobile data.

The terminal is only half of it. My agents build things I need to see: a demo, a proof of concept for a client, or just the app they're working on. But the box is locked down, and none of those apps is open to the internet.

So I use Cloudflare Tunnel. A small program on the box, cloudflared, opens a connection out to Cloudflare. The box never has to accept a connection in. When I open one of my apps in a browser, Cloudflare carries the request down that connection to the app on the box. In front of every app sits a Cloudflare Access login that only lets my email address through: I type my address, Cloudflare emails me a one-time code, and I'm in. To show a client a demo, I add their email address for that one app.

One rule makes it work: every app on the box listens on the box itself (localhost) and never on its public address. A tunnel doesn't close anything on its own. If an app listened on the public address, anyone could reach it around Cloudflare and around the login.

And it's cheap. The KVM 8 (8 vCPU / 32 GB, Hostinger's biggest VPS) is $25.99 a month on a two-year plan, $623.76 paid upfront, then $49.99 a month when it renews (Hostinger's US price today). Cloudflare Tunnel costs nothing, and Access is free for up to 50 users. That's the whole bill for a machine that runs my agents around the clock and shows me their work on my phone.

And all of a sudden, everything happened in the command line. There's no text editor on a VPS box. I never installed one, and I never needed to.

When something needs editing, which is rare, I ask the agent to do it. For the odd tiny change I use Vim. Even setting up an API key doesn't need an editor: I ask Claude for a script with a placeholder, paste it into the terminal, and replace the placeholder with the value.

It didn't happen overnight. It took three stages, and each one moved the editor further from the start of my day.

I don't miss the editor. Nothing is worse without it. You just have to rethink how you build software from the ground up.

More productive, no question

I'm far more productive, and it isn't even close. Two reasons.

No limit on the number of agents. On my laptop, memory decided how many agents I could run. On the server, I run as many as the work needs.

They work when I can't. They keep going while I sleep, while I'm interrupted, while I'm on the go. A typical moment: I'm shopping with my wife, she's off looking at dresses, and I sit down with a coffee and check on every agent from my phone. Nothing waits for me to be back at a desk.

My agents work while I sleep, and I check on them over a coffee.

And it's going further. I'm building my own software factory, where agents run for days on their own. On the server, days on their own is just another run. I've written about what a software factory is, and the philosophy I'm building mine on, in Stop building software. Build the software factory that builds it.

I review outcomes, not code

Here's the part that will upset people: I don't review the code my agents write. At all. I review outcomes. Reviewing code is a thing of the past.

Engineers were never paid to love code. They were paid for outcomes.

In my experience, there has always been friction between developers and the business. Business people don't like talking to developers, which is why companies keep a layer of middle management between them. What the business wants is outcomes. Too many engineers cared more about the craft of coding than about the outcome.

I used to be one of them. Everything changed when I became a business owner and started my own consultancy. There's a place for both, but it's becoming more and more critical that engineers, in software and in every other discipline, think in business outcomes and see the big picture. That doesn't mean they stop caring about the craft. It means their craft serves the business outcome. Honestly, I find that a constructive constraint for the craft, not a crippling one.

Agents force that change. Much of what a business cares about (performance, security, reliability) can be written down as checks that a machine can verify. So the engineer's new job is verifiability: turn each outcome into deterministic tests, unit and end-to-end, that prove it, and let the agent loop until they all pass. I wrote about that loop in How much freedom should an AI step get? That's why I don't need to read the code, or open an editor to read it in.

Passing tests are where it starts, not where it ends. The real test is whether the work meets the objective it was built for, once it's in front of customers. That's the heart of my software factory. How agents change the economics of verification has a post of its own: Verifiability is your next AI investment. TDD is just the start.

Editors will die. Prime suspect: software factories

Here's what the numbers say today:

Seven developers in ten still use VS Code. Editors aren't dead yet. But my bet is that they die, and VS Code with them: nobody will open one to write software. I may be wrong. I'd rather be clearly wrong than vaguely right.

Here's why I'm confident. Software factories are where code is going to be written, and a software factory doesn't need a text editor. It's an environment you build for your agents: it takes a request all the way to a business outcome in production, and nobody opens a file along the way. Within the next 12 months, I expect software factories to produce 70% or more of all the source code written in the world, maybe more.

It's the same thing that already happened to web traffic. In June, Cloudflare's CEO, Matthew Prince, said bots had passed humans on the internet for the first time, sooner than he had predicted. Agents don't browse like people, and they don't write code like people either. If you want to think about where this goes, I highly recommend Cloudflare CEO Matthew Prince and Nous CEO Dillon Rollnick on who owns the internet in the age of AI, a fascinating conversation on the future of the internet, from Union Square Ventures.

Microsoft is asleep

I dislike Microsoft's developer tools, and I've never hidden it. The one thing Microsoft does well today is buying companies. Good on them for buying GitHub in 2018, and npm through GitHub in 2020.

Copilot is junk. If you're still using it, stop wasting your time and go straight to Claude Code.

Microsoft has noticed something is moving. VS Code 1.145, released on 4 November, lets one session mix agents from different vendors: in its own example, Claude reviews a plan that a Copilot chat wrote (a preview feature, off by default). Copilot Fusion, generally available since 28 October, picks and orchestrates models for you instead of asking you to choose one. The strategy is to make the editor the host for every agent.

That's barking up the wrong tree. Packing more into VS Code only matters if people keep opening it, and I think they'll stop. Microsoft isn't adapting to the shift. It's decorating the room everyone is about to leave.

And GitHub?

Thank God they bought GitHub. It's still where code lives and where agents work: agents open more than a third of the pull requests on it. It's still my default today.

But GitHub is under attack:

• Its own code. In May, an attacker reached GitHub's internal repositories through a poisoned VS Code extension on an employee's device. The attacker claims about 3,800 repositories, which GitHub called "directionally consistent" with its investigation.
• Its features, used against its users. On 1 September, attackers hid malicious commands in issues, pull requests and comments: 83 of them across 61 public repositories, most published within 43 minutes.
• Its uptime. On 17 August, pull requests, issues, webhooks and Actions were degraded for hours.
• Its competitors. The same day, Cursor launched Origin, its own git forge, in early beta: repositories, pull requests and GitHub sync, built for agents. And in October, Anthropic said it had moved its monorepo off Git to Scion, a version control system it built itself, after its daily merges went from about 2,500 in May to more than 30,000 in September. Anthropic says Scion isn't a replacement for Git in general, but its reason is the one that matters: Git assumes people working with a full copy on their own machine, and Scion assumes agents and jobs that are always online.

GitHub became GitHub because of the team that built it: customer-obsessed hackers who cared about developers' problems at the time. Not because of Microsoft. I've always thought Microsoft is bad at product, and I wouldn't be surprised if it answers this moment too slowly, or with the wrong thing.

And the ground under GitHub is moving too. A pull request isn't part of Git. It's a practice, a convention built on top of Git. I don't use pull requests any more. When agents write the code, who owns the code changes, and when ownership changes, the whole development lifecycle changes with it. I think the software factory is the new lifecycle: the work isn't done when the ticket closes, it's done when the outcome is met. Pull requests may well die, replaced by something we haven't named yet.

Strip that away, and GitHub's real advantage today is that it's free: unlimited repositories, private ones included, and private is what most of us need. The day someone offers something better that's also free, I expect an exodus. I have zero trust in Microsoft's ability to make that turn.

What would prove me wrong

• Editors become the agent's home. That's exactly where VS Code is heading. If the best agent experience ends up inside an editor, developers come back to it, and Microsoft's bet pays off.
• GitHub wakes up. If it moves faster than its challengers and reinvents what comes after the pull request, it keeps its place.
• One developer isn't the market. I work on my own projects and my clients' systems with agents every day. Most developers work inside teams, rules and tools they didn't choose, and those change slowly.
• VS Code becomes a software factory. This one would prove me wrong and right at the same time. I suspect it's coming, and very soon: Microsoft will try to turn VS Code into a software factory. VS Code is already a Frankenstein: a code editor with a plugin for everything, wired to Copilot and every other agent. It's already a mess, and a factory would be one more limb on the mutant. If people keep opening something called VS Code only to run a factory, then yes, VS Code survived. But the code editor didn't: the editing is the part that disappears. And if it works well, my guess is that Microsoft eventually ships it as a new product, with a new name that has nothing to do with VS Code. The first step, though, will be to try software factories inside VS Code.

My bet

Developers will spend their attention in the agent session and the software factory, where the work starts, and on the outcomes they prove, where the work ends. Engineering teams will stop building software, and build the software factories that build it. The editor becomes a place nobody visits. Microsoft owns the window everyone is walking away from, and it's busy redecorating it.