Man, it is a really weird time in the software world right now. Like a lot of people, when I picked up ChatGPT and Midjourney back in late 2022. It felt like wizardry.
For a year I experimented and experimented. I wrote about how Midjourney responds to emojis and even tried to see if I could write a children’s book with AI. The output here looks pretty rough today, but it was a revelation at the time. And the AI tools just kind of gradually got better and better.
One thing that never really clicked for me though was agentic coding. The vision is that you tell an LLM what you want, and it just goes off and executes it. It never really worked that well. It felt like the agents just kind of plowed through code and broke a bunch of stuff on the way to making the fix I wanted. Coding agents were more a curiosity, not something I could actually use.
However, this is all about to change. And it is going to change everything about how software is built.
Gas Town
I was drawn back by a post called Welcome to Gas Town by Steve Yegge.
Gas Town is part visionary, part performance art. It’s a Mad Max-themed fever dream that enables agents, managing agents, managing agents. There’s a mayor, rigs, polecats, a deacon and a refinery—and they all work together in this vast factory where vibes go in and code comes out.
There’s a lot of really clever ideas that somewhat mesh together. You’ve probably heard about the context window LLMs have. Essentially it’s their short-term memory. Once an agent uses about 20% of its context window, its intelligence drops off a cliff and it starts doing insane things like dropping databases.
Gas Town employs a trick to manage that issue. It assigns tasks to ephemeral “Polecat” agents. Polecats do a task and then disappear—basically removing the challenge of managing context windows.
Maggie Appleton articulated the value of Gas Town well:
We should take Yegge’s creation seriously not because it’s a serious, working tool for today’s developers (it isn’t). But because it’s a good piece of speculative design fiction that asks provocative questions and reveals the shape of constraints we’ll face as agentic coding systems mature and grow.
And so, intrigued by Gas Town, I decided to try vibe coding again.
The Vibes Are Real
I’m not the type to just dip my toes in. If I’m going to do something, I do a cannonball.
I’m always looking for the meta. What is the best strategy and who knows how to execute it? I read that Anthropic’s CTO keeps five agents running constantly and barely looks at the actual files. Either that’s marketing or there must be something there, or maybe both?
I read a bunch of articles and watched a ton of YouTube videos. YouTube was pretty wild. There are these videos with guys streaming 5-10 Claude Code terminals, blaring EDM (lol) and managing all these agents.
I’m kind of poking fun, but I actually learned a lot about setup from that BridgeMind channel. If you’re interested I might start with his videos about Warp and the OpenCode CLI.
I wanted to see how well these agents actually work, but I needed an easy entry point. A Chrome extension to restyle Hacker News seemed perfect. It’s been in my backlog for a while, and because it’s purely frontend, I knew I’d be comfortable judging the output.
And honestly it was. Initially I tried to one-shot the thing and—as I expected—that was a failure. But then I decided to slow down. I told the agent to scaffold a chrome extension to restyle Hacker News pages. It worked. And I kind of just broke up these tasks into smaller pieces. Tested them as I went.
And it kind of just worked.
Sure there were bugs. But I kind of just did what I do when I’m reviewing any engineers’ code. Inspect the DOM, look at the styles, look at the console and then give feedback. When the context window hit’s 20% I typically close that window and open a new chat. The OpenCode CLI even allows you to drop screenshots. It’s pretty wild.
Now, I don’t have a dozen agents running at the same time. I never felt the need to have more than two working. And I’m honestly not sure how UI work even gets shipped in Gas Town? My guess is Yegge is probably a lot less concerned with UX than I am.
With AI agents, the last mile—that final polish and detailing—will be critical. We already see this in Salesloft, where sellers review generative emails before sending. In design, it manifests as small UI tweaks. It will be something else for doctors and something else for mechanical engineers. But I think there is a real opportunity in refining how humans interact with the agent’s output, creating better loops for feedback and adjustment.
You know what? There’s something here though. I don’t really like the term “vibe coding.” And I know the concept is polarizing. But after tinkering with this stuff for a couple days—I think agents are about to change how we build software.
Product Engineer
The conception of what it looks like to make software is going to change pretty quickly.
The three major functions on a delivery team (or feature team) are engineering, design and product management. I’ve long thought we’re going to start to see more overlap in those functions. I’m even more sure of it now.
I think we’re going to start to see a hybrid role emerge—product engineer.
What does this mean delivery teams will look like in the future? I imagine they are either significantly smaller or significantly more productive. I’m not sure if QA is embedded into these teams the same way they are currently or if there is a separate team of—well, people managing QA agents. I have a lot of questions.
This also makes me think a lot about Ben Thompson’s theory on bundling and unbundling. From 2010 to 2015, companies quickly moved on the back of frameworks. First like Ruby on Rails and Bootstrap. Then on other technologies like Angular and React. The speed these frameworks provided caused an unbundling in the software world.
Point solutions were able to move fast and gather steam while slow incumbents either weren’t nimble enough or weren’t in a place to capitalize on the productivity provided by frameworks. Starting in 2016, that changed. Customers were overwhelmed with choices. Larger companies caught on and smaller ones consolidated into larger platforms.
I think that’s about to shift again. And probably this year.
David Cummings called it out in his newsletter this weekend. SaaS companies are about to see a massive wave of new competition. And it is going to happen extremely fast. The bar to build software has been lowered. A two-person team will soon be able to build what used to take a whole department.
This puts incumbent software companies in a pretty dangerous situation. Those that are not able to be nimble and go fast are going to be in real trouble. I think this is especially true in the consumer, SMB and mid-market segments. Enterprise software may have some buffer as customers of enterprise software are buying a process more than the software itself.
Closing
Honestly, it makes me a little nervous. The industry is going to change and everyone’s jobs are going to look a little different—product design included.
But as someone who got into software just because I wanted to make things, this is a dream come true. I’m seeing a glimpse of the vision I hoped for in 2022. It’s not just a toy anymore. It’s an unbelievable tool for builders.
If you’ve been ignoring AI tools because you think they are overhyped, or maybe don’t see how they fit into your workflow. I’d encourage you to give them another look.
It’s time.