The next article initially appeared on Drew Breunig’s weblog and is being republished right here with the writer’s permission.
In 1998, Eric S. Raymond printed the founding textual content of open supply software program improvement, The Cathedral and the Bazaar. In it, he detailed two strategies of constructing software program:
- The cathedral mannequin is rigorously deliberate, closed-source, and managed by an unique group of builders.
- The bazaar mannequin is open, clear, and community-driven.
The bazaar mannequin was enabled by the web, which allowed for distributed coordination and distribution. Extra folks may contribute code and share suggestions, yielding higher, safer software program. “Given sufficient eyeballs, all bugs are shallow,” Raymond wrote, coining Linus’s legislation.
The concepts crystallized in The Cathedral and the Bazaar helped kick off a quarter-century of open supply innovation and dominance.
However simply because the web made communication low-cost and birthed the bazaar, AI is making code low-cost and kicking off a brand new period stuffed with idiosyncratic, sprawling, cobbled-together software program.
Meet the third mannequin: The Winchester Thriller Home.
The Winchester Thriller Home
Situated lower than 10 miles southeast from the Laptop Historical past Museum, the Winchester Thriller Home is an architectural oddity.
Following the dying of her husband and mother-in-law, Sarah Winchester managed a fortune. Her shares within the Winchester Repeating Arms Firm, and the dividends they threw off, made it so Sarah couldn’t solely dwell in consolation however pursue no matter ardour she desired. That zeal was structure.
Sarah didn’t construct her mansion to deal with ghosts1; she constructed her mansion as a result of she preferred structure. With no license, no formal coaching, in an period when girls (even very wealthy girls) didn’t have a path to practising structure, Sarah centered on her own residence. She made up for her lack of license with ardour and successfully limitless funds.
Sarah constructed what she wished. “At its largest the home had ~500 rooms.” Right now it has roughly 160 rooms, 2,000 doorways, 10,000 home windows, 47 stairways, 47 fireplaces, 13 bogs, and 6 kitchens. Carved wooden drapes the partitions and ceilings. Stained glass is in all places. Tasks have been deliberate, accomplished, deserted, torn down, and rebuilt.
It was something however aimless. And sensible improvements ran all through, together with push-button gasoline lighting, an early intercom system, steam heating, and indoor gardens. The eccentricities that amuse right now’s guests have been principally sensible lodging for Sarah’s well being (stairways with very small steps), useful designs not used (lure doorways in greenhouses to route extra water), or fast fixes to wreck from the 1906 earthquake.
Winchester handed in 1922. 9 months later, the home grew to become a vacationer attraction.
Right now, many programmers are Sarah Winchester.
What occurs when code is reasonable
We aren’t as wealthy as Sarah Winchester, however when code is that this low-cost, we don’t have to be.
Jodan Alberts illustrated this just lately, accumulating and visualizing information detailing public GitHub commits attributed to Claude Code. That’s his information within the chart above, with Claude seeming to solely speed up by means of March.2
It’s arduous to get a deal with on particular person utilization although, so I went looking for a proxy and landed on the chart under:
After Opus 4.5 and up to date work enabling Agent Groups, the typical internet traces added by Claude per commit is now clean and regular at 1,000 traces of code per commit.3
1,000 traces of code per commit is ~2 magnitudes increased than what a human programmer writes per day.
For those who seek for human benchmarks, you’ll discover many citing Fred Brooks’s The Legendary Man Month whereas claiming engineer would possibly write 10 cumulative traces of code per day.4 For those who additional discover, you’ll discover numbers increased than 10 cited, however usually lower than 100.
Right here’s anecdote from antirez on a Hacker Information thread discussing the Brooks “quote”:
I did some trivial math. Redis consists of 100k traces of code, I wrote at the very least 70k of that in 10 years. I by no means work greater than 5 days per week and I take 1 month of holidays yearly, so assuming I work 22 days each month for 11 months:
70000/(22 x 11 x 10) = ~29 LOC / day
Which isn’t too removed from 10. There are days the place I write 300-500 LOC, however I suppose that plenty of work went into rewriting stuff and fixing bugs, so I rewrote the identical traces many times over the course of years, however but I feel that this ought to be taken under consideration, so the Legendary Man Month e-book is certainly fairly correct.
Six years after this remark, Claude is pushing 1,000 traces of code per commit.
So what can we do with all this low-cost code?
Sadly, all the pieces else stays roughly the identical value and roughly the identical velocity. Suggestions hasn’t gotten cheaper; the “eyeballs” that guided the software program developed by the bazaar haven’t caught as much as AI.
There is just one supply of suggestions that strikes on the velocity of AI-generated code: your self. You’re there to immediate, you’re there to overview. You don’t have to recruit testers, run surveys, or handle design companions. You simply construct what you need and use what you construct.
And that’s what many builders are doing with low-cost code: constructing idiosyncratic instruments for ourselves, guided by our passions, style, and wishes.
Sound acquainted?
Welcome to the thriller home
Steve Yegge’s Fuel City is a Winchester Thriller Home. It’s extremely idiosyncratic and sprawling, wealthy with metaphors and hacks. It’s the right device for Steve.
Jeffrey Emanuel’s Agent Flywheel is a Winchester Thriller Home. A big subset of tokenmaxxers resolve they should rebuild their dependencies in Rust; Jeff is one such instance. His “FrankenSuite” contains Rust rewrites of SQLite, Node.js, btrfs, Redis, pandas, NumPy, JAX, and Torch.
Philip Zeyliger famous the sample final week, writing, “Everyone seems to be constructing a software program manufacturing facility.” However it goes past software program. Gary Tan’s private AI committee gstack is a Winchester Thriller Home constructed principally from Markdown.
In every single place you look, there are Winchester Thriller Homes.
Every Winchester Thriller Home is idiosyncratic. They’re extremely customized. The tightly coupled suggestions loop between the coding agent and the consumer yields software program that displays the developer’s wishes. They often lack documentation. To outsiders, they’re inscrutable.
Winchester Thriller Homes are sprawling. Guided by the wants of the developer, these instruments are likely to unfold out, always annexing territory within the type of new capabilities and new repositories. Work is sort of all the time additive. Code is added when it’s wanted, bugs are patched in place, and numerous appendages stay. There’s little incentive to prune when code is free.
And constructing a Winchester Thriller Home ought to be enjoyable. Coding brokers flip all the pieces right into a aspect quest, and we eagerly take part. Constructing the right workflow is a ardour for a lot of devs, so we hold pushing.
Winchester Thriller Homes are idiosyncratic, sprawling, and enjoyable. However does this imply we’re abandoning the bazaar?
What occurs to the bazaar?
What occurs after we all are likely to our thriller homes? When our free time is spent constructing instruments only for ourselves, will we cease engaged on shared tasks? Will we abandon the bazaar?
Most likely not. The bazaar is packed proper now, however not in a great way.
Code is reasonable, so individuals are slamming open supply repositories with agent-written contributions, in an try to pad their résumés or manifest their pet options. Daniel Stenberg ended bug bounties for curl after a deluge of poor submissions sapped reviewer bandwidth. It’s gotten so dangerous, GitHub just lately added a function to disable pull request contributions.
Anecdotally, I’m seeing good contributions choose up as effectively. They’re simply drowned out by the slop. For what it’s value, curl commits are dramatically up within the agentic period. And other people are sharing what they construct. A current evaluation by Dumky exhibits packages and repos rising within the final quarter.
There’s loads of finances for each thriller homes and the bazaar when code is this low-cost. The brand new problem is creating techniques and processes for managing the deluge. We don’t want eyeballs to search out bugs in the software program; we’d like eyeballs to search out bugs earlier than they attain the software program.
In some ways that is the inverse of the bazaar mannequin period. The web made suggestions and communal coordination quicker, simpler, and cheaper. The bazaar mannequin has a excessive throughput of suggestions (many eyeballs) however comparatively excessive latency for modifications (file a problem, focus on, submit a PR, look forward to overview, and so forth.).
Coding brokers, then again, make implementation quicker whereas suggestions and coordination are unchanged. The Winchester Thriller Home mannequin sidesteps this by collapsing the suggestions loop into one particular person: Latency is close to zero, however throughput is simply you. The bazaar, outlined by communal work, can’t undertake this hack. Coding brokers within the bazaar create a large number: implementation at machine velocity hitting coordination infrastructure constructed for human velocity. Which is why maintainers really feel like they’re drowning.
We want new instruments, expertise, and conventions.
Classes from the thriller home
Coding brokers have dropped the price of code so dramatically we’re getting into a brand new period of software program improvement, the primary change of this magnitude for the reason that web kicked off open supply software program. Change arrived rapidly, and it’s not slowing down. However in reviewing the Winchester Thriller Home framework, I feel we will take away just a few classes.
Lesson 1: The bazaar and Winchester Thriller Homes can coexist.
When itemizing instance Winchester Thriller Homes, I didn’t point out OpenClaw, despite the fact that it’s the defining instance. I saved it for right here as a result of it properly illustrates how Winchester Thriller Homes and the bazaar can coexist.
OpenClaw is extremely modular and locations few limitations on the consumer. It integrates 25 totally different chat and notification techniques, plugs into most inference finish factors, and is constructed on the exceptionally versatile pi agent toolkit. This keen flexibility was embraced early—safety and information protections be damned—however since its exponential adoption Peter Steinberger and the group have been steadily pushing enhancements and fixes.
And like different breakout open supply tasks of yore, the ecosystem is adopting the very best concepts and mitigating the worst features of OpenClaw. Numerous alternate “claw” tasks have emerged. (There’s NanoClaw, NullClaw, ZeroClaw, and extra!) Firms have launched providers to make claws simple or safer. Cloudflare launched Moltworker to make deploy simple, Nvidia shipped NemoClaw with a safety focus, and Claude retains including claw-like options to its desktop app.
Lesson 2: Don’t promote the enjoyable stuff.
One motive OpenClaw works so effectively within the bazaar is that it’s a basis for private instruments. Out of the field, a claw simply sits there. It’s as much as the consumer to find out what it does and the way it does it, leveraging the connections and infrastructure OpenClaw gives. OpenClaw lets much less skilled builders spin up their very own Winchester Thriller Homes, whereas skilled devs get to leverage a lot of the frequent integrations and techniques OpenClaw gives. Peter and group have finished an amazing job drawing a line between the frequent core (what the bazaar works on) and what they go away as much as the consumer: The boring, essential stuff is the job of the commons.
Considering again to Sarah Winchester and her idiosyncratic, sprawling mansion, we see the identical sample. Sarah employed distributors! She used off-the-shelf elements! Her bathtubs, bathrooms, taps, and plumbing weren’t crafted on web site.
The boring stuff, the arduous bits, or the issues which have disastrous failure modes are the issues we should always collaborate on or make use of specialists to deal with. (Come to suppose, plumbing checks all three containers). That is the chance for open supply software program, dev instruments, and software program corporations.
Don’t attempt to promote builders the stuff that’s enjoyable, the stuff they need to construct. Promote them the stuff they keep away from or don’t wish to take accountability for. Sarah Winchester didn’t rent metalworkers to craft the pipes for her plumbing, however she did rent craftspeople to create a whole lot of stained-glass home windows to her specs.
Lesson 3: The bounds of code are communication.
OpenClaw exhibits the bazaar stays related but additionally highlights the issues dealing with open supply within the agentic period. Proper now, there are 1,173 open pull requests and 1,884 new points on the OpenClaw repo.
There may be extra code and extra tasks than we may ever overview. The problem now, for open supply maintainers and customers, is sifting by means of all of it. How do we discover the novel concepts that everybody ought to undertake and borrow?
OpenClaw is without doubt one of the successes, one thing we all observed. And for it, the issue is processing the suggestions. For the tasks we’ll by no means discover, those misplaced within the deluge, their drawback is lack of suggestions. You both discover consideration and drown in contributions or drown within the ocean of repos and by no means hear a factor.
The web made coordination low-cost and gave us the bazaar. Coding brokers made implementation low-cost and gave us the Winchester Thriller Home. What we’re lacking are the instruments and conventions that make consideration low-cost, that allow maintainers take in contributions at machine velocity and let good concepts floor among the many noise. Till we determine this out, the bazaar will hold getting louder with out getting smarter, and the very best concepts in our thriller homes might be forgotten as soon as we cease sustaining them.
