Witnesses, polecats and nuxes. Surpringly this is not Mad Max, this it the cutting edge of agent driven development. Even more surprisingly I only had to make one of those words up.
Gas Town is Steve Yegge's... agent orchestrator. I hesitate as the way he talks about it can be quite full on:
The industry is an embarrassing little kid’s soccer team chasing the 2025 CLI form factor of Claude Code, rather than building what’s next. I went ahead and built what’s next. First, I predicted it [...]
Gas Town is the tip of a deep iceberg. Gas Town itself may not live longer than 12 months, but the bones of Gas Town — the MEOW stack — may live on for several years to come. It feels like more of a discovery than an invention.
And that's missing the part where you have to be at stage seven on his AI journey chart to even think about using it. This trend in the all-in-on-AI community has been analysed better, and with a lot more grace than I could manage, by Armin Ronacher.
Putting the language aside the idea that you can graft all these janky things together feels like the antithesis of good software engineering. Fragile, wasteful and chaotic it's not exactly the measured, thoughful profession we like to think we are a part of.
Intentionality
Gas Town, the people who use it and things like it know what they're doing. They're true believers in AI and that it is the future of Software Engineering. The fact that the resulting code is a mess is an intentional trade-off they've made, not an accident.
Most of us are not there - we're in the genteel suburbs. If we use AI perhaps it's as autocomplete, a rubber duck or to write laborious scripts and boilerplate. Maybe the stage four-ers amongst us let the agent run wild but then do code review. We want - or are being encouraged by our employers - to improve our productivity using these tools without sacrificing on quality. More and more of us are going to move into the fourth stage.
We're intending to do both what feels right (high quality) and what has been asked of us (high productivity). What was that saying about the road to hell?
Straddling Two Horses
AI is often compared to the Industrial Revolution and the invention of the loom, which took us from slow but high quality craftmanship to mass produced garments overseen by machine operators.
For those living outside of Gas Town, but not in the backwards marshes, we seem to think that we can - and everyone wants - to keep quality high. This is despite that there is more and more evidence that quality, usefulness or even just having a product that basically functions the way it has for decades isn't that important when weighed against shareholder value. AI is just going to accelerate this process.
We have also missed the fact that operators didn't need to be skilled weavers. We think the natural order of things is that with AI we will need to be a skilled software engineer. Even if the incentives were aligned to quality, I am not so sure "agent as a coder, SWE as a reviewer" works that way. Yes, it robs junior engineers of learning opportunities but it also atrophies senior's reviewing skills.
Could you teach someone who doesn't code to be a good reviewer? Then how do we expect to keep our skill in reviewing if we're not coding? The best reviews I've had are from the people who first wrote the code I am changing because the act of coding has given them a mental model of what it does. That's not something you get by a scan of some agent's output.
(I Don't Want To Go To) Gas Town
The only thing worse than living in Gas Town is going there by accident. By accepting more and more AI input into our jobs and our lives I think - although we might have the best of intentions - that's where we're going. A mixture of the pressures that have driven enshittification and the degredation of our skills will buy us a one way ticket there.
Well I don't want to go. I don't want products that are a mess with bugs that no human understands. I don't want those bugs explained by chirpy, sycopahntic prose written by a blurry fox mayor. I don't think it's fair that work people put out in good faith as free software that needed attribution has been hoovered up to enrich the already obscenely rich. And selfishly: I want to weave, I don't want to spend half my time risking an arm by sticking it into a machine and the other half sat around bored.
I don't know what the answer is. A profession that can't get maintenance work (tech debt) scheduled doesn't seem well placed to carving out time to keep our skills up, nevermind avoiding LLMs entirely. An industry that has created algorithms that amplify hate speech - possibly fueling a genocide - is unlikely to care about a low quality product.
We should still try.