The Silver Bullet Is Here
In 1986, Fred Brooks wrote “No Silver Bullet”, later collected in his seminal work, The Mythical Man-Month. His thesis was that no single development, in either technology or management technique, would deliver even a tenfold improvement in software productivity within a decade. Well…
Brooks also wrote:
The hardest single part of building a software system is deciding precisely what to build.
You would be a fool to think that AI won’t be able to do this better than a human can. Sure, a while ago you could argue that the human had all the business context, and had to stay in the loop. LLMs can have way more context than you. Is AI missing some context about something? Write it down, or the LLM will!
If there’s anything I’ve learned in the last ten years, it’s to never say “computers will never be able to do X.”
When I was a kid, I grew up reading about Deep Blue beating Garry Kasparov. I remember thinking to myself: “Why would anyone think a computer couldn’t beat a human at chess?”
Then I learned about Go. I thought to myself: “Now there’s a game that computers will never be better at.”
When AlphaGo beat Lee Sedol in 2016, that was the first time I really started to think that maybe there isn’t anything computers can’t do, even in principle.
Today, in 2026, I’m here to tell you that LLMs really are a silver bullet. If you haven’t been living under a rock for the last year, you know LLMs have dramatically changed how software is written.
Look at agent swarms. Steve Yegge’s Gas Town runs 20 to 30 coding agents in parallel on the same codebase. Anthropic had 16 parallel Claudes build a C compiler from scratch in two weeks, one good enough to compile the Linux kernel. Cursor ran hundreds of agents for a week and built a web browser, over a million lines of code. Even plain Claude Code will happily fan out subagents to work in parallel.
Still don’t think you can 10x yourself?
When I first read about Gas Town, I thought it was super silly: a farce of a software development methodology.
I still think it’s very silly. But it is on to something.
The future is definitely not dozens of agents stepping on each other’s toes and wasting tons and tons of tokens. But the future is multiple agents working simultaneously, in a controlled way, under one logical leader.
The ranks of camp skeptic have dwindled as LLMs become more and more proficient, and able to reason at higher and higher levels.
It’s inevitable that soon, humans won’t be writing code at all (except for fun).
The calculus for what it takes to write and maintain software has changed. Our heuristics need recalibration. The old rules of thumb about what code costs to write, integrate, and maintain no longer hold.
The next evolution will be taking the human out of the loop entirely.
The Cathedral and the Bazaar
In his famous 1997 essay “The Cathedral and the Bazaar”, Eric S. Raymond describes two different ways to build software. In the cathedral, a small group of experts works in private and hands down finished releases. In the bazaar (the Linux model), you release early and often, delegate everything you can, and treat your users as co-developers. Raymond argued that the bazaar produces better software, faster.
Raymond’s very first lesson is that “every good work of software starts by scratching a developer’s personal itch.” The bazaar is what happens when lots of people have the same itch. Nobody can afford to build and maintain an entire operating system for themselves, so we pool our effort and all share the result.
Do you think this is still true? Is “developer” time scarce and expensive now?
What happens when developer time becomes nearly free?
The Hyper-Cathedral: Forks Go Private, Open Source Goes Darker
Let’s look at the Linux kernel forks for a bit.
Google runs its production fleet on Prodkernel, a private fork of Linux carrying about 9,000 patches. Every couple of years, Google rebases all of those patches onto a new upstream kernel. This is a kind of insanely painful thing that companies only do when they have an enormous bank of engineering power and a really good justification for doing it. Some people call it tech debt! Prodkernel technically isn’t a cathedral, because it’s not open source. Android has the same story, with vendor kernels forked from forks of forks: Google forks each kernel into the Android Common Kernel, SoC vendors fork that for each chip, and phone makers fork it again for each device. This got bad enough that Google built the Generic Kernel Image to pull everyone back toward one kernel.
Eventually the cost gets too high and the organization decides to go back to the bazaar. But what if it was nearly free to maintain a Linux fork?
In the hyper-cathedral, everyone with access to LLMs collectively decides that developing together simply isn’t worth it, when an AI can maintain a fork tailored to only your own needs. The key attribute of a hyper-cathedral is that it is autonomous. LLMs just take care of it.
With advanced AI, there’s simply no incentive to contribute back. Even if you did, it would be of little value. Why spend your tokens making pull requests and waiting on reviews, when you could just charge ahead?
And plenty of people won’t bother publishing at all. Prodkernel-style private forks will become the norm.
The Hyper-Bazaar: Collaborative Open Source Goes Superhuman
In the hyper-bazaar, we take humans out of the loop, and LLMs are solely responsible for maintaining projects.
OpenClaw is the closest thing to this today that I could find. Peter Steinberger started it as a weekend project in November 2025, and by August 2026 it had over 80,000 commits. Steinberger himself runs around 100 agents that write code, review pull requests, and hunt for bugs. He calls pull requests “prompt requests,” and says he ships code he doesn’t read.
I would call this a hyper-bazaar.
Picture an LLM-run fork of a popular project. Pull requests come from corporations and individuals with their own LLMs and tokens to spare. Caretaker LLMs review all the incoming code and merge it at speed.
LLM-run hyper-forks will become canonical if humans can’t keep up on the original.
The hyper-bazaar will scale to the same kind of high-frequency continuous deployment we already see inside enterprises. Daily releases. Agents everywhere pulling the latest code on every build. A continuous stream of fixes, merged as fast as they arrive.
I really hope LLMs will also let us keep multiple stable branches. Historically, that has been expensive.1 With LLMs, we could theoretically keep many major versions up to date at the same time. Please help us usher in a new era of LTS!
With LLM help, why couldn’t there be a maintained 2.4 Linux kernel?
Linux 2.4 came out in 2001. Willy Tarreau kept it alive with security fixes for years, but its last release was 2.4.37.11 in December 2010, and he called it quits at the end of 2011.
Wouldn’t it be really cool if, in 2027, the 2.4 kernel were revived, with continued support and releases, and no human in the loop to get it done? The first LLM-maintained, hyper-bazaar Linux branch?
The downside to the hyper-bazaar is the sheer chaos of it all. I’m mostly worried about xz-style malicious pull requests. Then again, humans couldn’t detect that one either.2 I don’t know whether LLMs are more or less susceptible to that style of “social” engineering. It just seems like LLMs might fall for some sort of “I need this PR merged or my mother will die” style trickery?
Conclusion
If you work in software engineering and you still have a job, you’re going to have to make a choice. We’ve always had to choose between building and buying (although you are always building, and always buying). Now the choice is between building your own private cathedral and participating in the hyper-bazaar.
Except it won’t really be you doing the building. It’ll be you, instructing your agents on what to do.
The hyper-cathedral will serve your needs well, but it will probably be a local maximum.
The hyper-bazaar will be chaotic. We’re already getting a taste of this in the industry. Both Anthropic and OpenAI have pointed their models at open source and turned up hundreds of real vulnerabilities, some of which had been hiding for decades. It remains to be seen whether the hyper-bazaar of open source will be more secure in the long run than all those private cathedrals.
Heck, I’m going to start my own hyper-cathedral with my port of Android to my tiny Palm phone. It’s a case where I can automate the rebasing, so a super old phone gets the most up-to-date version of Android. There’s no real reason for me to try to upstream my changes for a phone with only a handful of users, but the effort for me to stay up to date is minimal.
I’ll let somebody else run a hyper-bazaar, like OpenClaw.
-
In 2023, the Linux kernel cut long-term support from six years to two. Maintaining a Linux kernel for years is a costly endeavor, at least in human terms. ↩︎
-
The xz backdoor was planted by a contributor who spent about two years earning the maintainer’s trust. It was only caught by luck, when Andres Freund noticed SSH logins taking half a second longer than they should. Veritasium made a documentary about the whole saga: “The Internet Was Weeks Away From Disaster and No One Knew”. ↩︎
Comment via email